Data on connections between the areas of the cerebral cortex andnuclei of the thalamus are too complicated to analyse with nakedintuition. Indeed, the complexity of connection data is one of themajor challenges facing neuroanatomy. Recently, systematicmethods have been developed and applied to the analysis of theconnectivity in the cerebral cortex. These approaches have shedlight on the gross organization of the cortical network, have made itpossible to test systematically theories of cortical organization, andhave guided new electrophysiological studies. This paper extendsthe approach to investigate the organization of the entire cortico-thalamic network. An extensive collation of connection tracingstudies revealed ~1500 extrinsic connections between the corticalareas and thalamic nuclei of the cat cerebral hemisphere. Around850 connections linked 53 cortical areas with each other, andaround 650 connections linked the cortical areas with 42 thalamicnuclei. Non-metric multidimensional scaling, optimal set analysisand non-parametric cluster analysis were used to study globalconnectivity and the ‘place’ of individual structures within theoverall scheme. Thalamic nuclei and cortical areas were in intimateconnectional association. Connectivity defined four major thalamo-cortical systems. These included three broadly hierarchical sensoryor sensory/motor systems (visual and auditory systems and a singlesystem containing both somatosensory and motor structures). Thehighest stations of these sensory/motor systems were associatedwith a fourth processing system composed of prefrontal, cingulate,insular and parahippocampal cortex and associated thalamic nuclei(the ‘fronto-limbic system’). The association between fronto-limbicand somato-motor systems was particularly close.

IntroductionEmpirical neuroanatomy has had tremendous success in tracing

the connections that link different regions of the brain.

However, individual connection tracing studies focus on the

connections of a few brain structures in a few individuals.

Therefore, each study describes only a small piece of the

connectional architecture of the whole brain. Trying to piece

together such data for an accurate overview of brain organiza-

tion is like trying to solve an intricate jigsaw puzzle. There are

many complimentary approaches to help put the jigsaw

together. High-quality anatomical studies give us the right pieces.

Physiological and behavioural data help show how the pieces fit

together. We also think that systematic collation and analysis of

connection data are valuable tools to help assemble the puzzle

and describe the picture (Young, 1992; Hilgetag et al., 1996;

Scannell, 1997).

Preliminary attempts to collate neuronal connection data have

already been published for the extrinsic connections linking

areas in the cerebral cortices of the cat (Scannell and Young,

1993; Scannell et al., 1995; Scannell, 1997) and the macaque

(Pandya and Yeterian, 1985; Felleman and Van Essen, 1991;

Young, 1992, 1993). Systematic analyses of the data show that

the cortical networks of cats and macaques contain three

sensory systems and a fourth system, termed the ‘fronto-limbic’

system, composed of prefrontal, insular, cingulate, perirhinal,

retrosplenial and entorhinal cortex and the hippocampal

formation (Scannell and Young, 1993; Young, 1993; Scannell et

al., 1995; Young et al., 1995a,b). The sensory systems all exhibit

a degree of hierarchy, both in terms of laminar origin and

termination patterns (Felleman and Van Essen, 1991; Young,

1992; Scannell et al., 1995) and because there is a topological

progression with some areas being connectionally close to

peripheral input and others being distant from peripheral input.

Systematic collation and data analysis have also shown that the

simple local wiring rules can account for much of the pattern of

cortico-cortical connections in the cat and macaque (Young,

1992; Cherniak, 1994; Scannell et al., 1995; Scannell, 1997).

They have shown that sulci and gyri are arranged in a way that

efficiently reduces the volume of the connections in the extrin-

sic cortico-cortical network (Scannell, 1996, 1997; Van Essen,

1997). In addition, systematic analysis of cortical connectivity

has had success in guiding electrophysiological investigations of

the cat visual system (Scannell et al., 1996), and in providing

simple network models that account for classical strychnine

neuronographic data (Hilgetag et al., 1997a).

It is important to extend the systematic approach to include

the thalamus. First, almost all neural signals from the sensory and

motor periphery reach the cerebral cortex via the thalamus (e.g.

Le Gros Clark, 1932; Jones, 1985). Second, cortical and thalamic

processing are closely co-ordinated (e.g. Le Gros Clark, 1942;

Kato, 1990; Guillery, 1995; Contreras et al., 1996; Crick and

Koch, 1998). Consequently, an undue emphasis on the cortical

networks ignores much of the machinery that may be

responsible for sensory and motor processing (Guillery, 1995).

This paper considers the extrinsic connections of both the

thalamus and cortex of the cat.

Researchers seeking large-scale principles of neuronal

organization face, in addition to the ‘jigsaw puzzle’ problem,

sampling bias and random sampling error in the primary

neuroanatomical literature. There is a strong bias towards a small

number of model structures (e.g. areas 17, 18 and LGN in the

case of the cat) for studies on laminar origin and termination

patterns, cellular targets, synaptic morphology, physiological

function, interindividual variability or development of connec-

tions. It is not clear if these structures are representative of the

rest of the brain. In contrast, systematic collation, while not

free from sampling bias, can yield relatively comprehensive

catalogues of the gross cortical connections (Young, 1993;

Scannell and Young, 1993; Scannell et al., 1995; Burns, 1997;

Scannell et al., 1997, and this paper). Therefore, although

lacking much of the richness and detail of the original reports,

the approach presented here is unusual in its breadth. We hope

that it can compliment the more traditional approaches to

neuroanatomy, upon which it is entirely dependent for data.

Cerebral Cortex Apr/May 1999;9:277–299; 1047–3211/99/$4.00

The Connectional Organization of theCortico-thalamic System of the Cat

J.W. Scannell, G.A.P.C. Burns, C.C. Hilgetag, M.A. O’Neil1 and

M.P. Young

Neural Systems Group, Psychology Department, Ridley

Building, University of Newcastle, Newcastle upon Tyne NE1

7RU and 1The Bee Systematics and Biology Unit, Hope

Entomological Collections, University Museum, Oxford

University, Parks Road, Oxford OX1, UK

We are currently extending our collation methods (e.g.

http://riposte.usc.edu/NeuroScholar) to emphasize available

data on laminar origin and termination patterns, cellular targets,

synaptic morphology, physiological function, interindividual

variability and development. We intend to apply systematic

analysis to these aspects of connectivity as and when data

become available for a wider range of connections.

Before systematically analysing connection data, it is neces-

sary to produce an accurate catalogue of connections. This step

presents major practical problems. The primary data differ in,

for example, individual animals, quality, method, anatomist,

parcellation and nomenclature. There is considerable variability

in the cortical connectivity of individuals, though less so in

thalamic connectivity (MacNeil et al., 1997). Parcellation is

difficult, especially in brain regions that lack systematic maps of

the sensory periphery (Colby and Duhamel, 1991). Further-

more, connection tracing is rarely quantitative, ‘interesting’

parts of the brain have been studied more than ‘boring’ parts,

and connections are not all the same. Therefore, it is necessary to

exercise considerable judgement at this stage of the study. These

problems notwithstanding, systematic collation is the only way

to produce an explicit summary of known connectivity. The first

part of this paper presents such a collation of the connectional

neuroanatomy of the cat thalamo-cortico-cortical network.

The second part of the paper presents the results of three

formal analyses that summarize the thalamo-cortical system’s

connectional architecture in a relatively compact and com-

prehensible way. The analyses illustrate the ‘place’ of individual

structures in the global connectional scheme and show global

features of the organization of the system that are not apparent

on casual inspection of the primary connection data. We hope

that both the collation and the analyses will guide new empirical

research and test organizational models of the thalamo-cortical

network as they have already done in the cortical network

(Scannell et al., 1996; Hilgetag et al., 1996; Scannell, 1997).

Materials and MethodsFull details of cortical and thalamic parcellation, connections and

citations of the original connection and mapping studies are published at

http://www.psychology.ncl.ac.uk/jack/cor_thal.html.

Table 1 gives a list of abbreviations used.

Cortical Parcellation

The cortical parcellation used in this study was based on Reinoso-Suarez

(1984), adapted by Scannell et al. (1995). Parcellation of the prefrontal

cortex and cortex on the medial side of the hemisphere (sometimes

termed ‘limbic’ cortex) was modified from Scannell et al. (1995) to bring

it into line with the schemes of Musil and Olson (1988a,b, 1991), Olson

and Jeffers (1988), and Olson and Musil (1992). All prefrontal cortex on

the lateral side of the hemisphere has been assigned to the area PFCl.

Areas ALG, SSF, SVA, DP, Amyg and 5m of the parcellation of Scannell et

al. (1995) have been omitted from this study because of a lack of data on

their connections with the thalamus or because of our uncertainty over

their identity or existence. Areas 4g and 4 of Scannell et al. (1995) have

been retained, but are probably cytoarchitectonic subregions of a single

primary motor area. Similarly, areas 3b, 1 and 2 probably constitute a

single primary somatosensory area, SI (Felleman et al., 1983). The

auditory cortical area VP is termed ‘VPc’ to distinguish it from the

ventroposterior nucleus of the thalamus.

Thalamic Parcellation

Parcellation of the thalamus was based on the atlas of Berman and Jones

(1982). In some thalamic regions (e.g. the lateral posterior complex, the

medial geniculate body, the ventrobasal complex) Berman and Jones’

terminology and parcellation were modified to conform to more recent

anatomical and physiological studies. The scheme used in this paper, and

its relation to Berman and Jones (1982) and to other recent studies, is

outlined in the Appendix.

The Appendix shows that there is a lack of consistency in

the nomenclature and parcellation used in different studies of certain

thalamic regions. Similar problems apply in the cortex (Scannell et al.,

Table 1Anatomical abbreviations

1 area 1 LGLI medial interlaminar nucleus17 area 17 LIc caudal region of the lateral

intermediate nucleus18 area 18 Llo oral region of the lateral intermediate

nucleus19 area 19 LM-Sg lateral medial-suprageniculate complex2 area 2 LPI lateral nucleus of the lateral posterior

complex20a area 20a LPm medial nucleus of the lateral posterior

complex20b area 20b MD mediodorsal nucleus21a area 21a MGDd dorsal nucleus of the medial geniculate21b area21b MGDdd deep dorsal nucleus of the medial

geniculate35 area 35 MGDds superficial dorsal nucleus of the medial

geniculate36 area 36 MGM magnocellular division of the medial

geniculate3a area 3a MGV ventral division of the medial geniculate3b area 3b MGvl ventrolateral nucleus of the medial

geniculate4 areas 4f, 4sf and 4d MV-RE medioventral-reuniens nucleus4g area 4γ P posteror auditory field5Al lateral area 5A PAC paracentral nucleus5Am medial area 5A PARA anterior paraventricular nucleus5Bl lateral area 5B PARP posterior paraventricular nucleus5Bm medial area 5B PAT parataenial nucleus61 lateral area 6 PF parafasicular nucleus6m medial area 6 PFCI lateral prefrontal cortex7 area 7 PFCMd dorsal medial prefrontal cortexAAF anterior auditory field PFCMil infralimbic medial prefrontal cortexAD anterodorsal nucleus PLLS posterolateral lateral suprasylvian areaAES anterior ectosylvian sulcus PMLS posteromedial lateral suprasylvian areaAI primary auditory field POi intermediate region of the posterior

complexAII secondary auditory field POl lateral region of the posterior complexALLS anterolateral lateral suprasylvian

areaPOm medial region of the posterior complex

AM anteromedial nucleus PS posterior suprasylvian areaAMLS anteromedial lateral

suprasylvian areapSb presubiculum, parasubiculum and

postsubicular cortexAV anteroventral nucleus PUL pulvinarCGa anterior cingulate cortex RH rhomboid nucleusCGp posterior cingulate cortex RS retrosplenial areaCLN central lateral nucleus Sb subiculumCM centre median nucleus SG suprageniculate nucleusCMN central medial nucleus SlI second somatosensory areaDLS dorsolateral suprasylvian area SlV fourth somatosensory areaEnr entorhinal cortex SSAi inner (deep) suprasylvian sulcal region

of area 5EPp posteror part of posterior

ectosylvian gyrusSSAo outer suprasylvian sulcal region of

area 5Hipp hippocampus proper SUM submedian nucleusIa agranular insula Tem temporal auditory fieldIg granular insula VA-VL ventroanterior-ventrolateral complexLD lateral dorsal nucleus VLS ventrolateral suprasylvian areaLGLA A laminae of the lateral

geniculateVMB basal ventromedial nucleus

LGLC magnocellular C laminae of thelateral geniculate

VMP principal ventromedial nucleus

LGLC 1–3 parvocellular C laminae of thelateral geniculate

VPc ventroposterior auditory field

LGLGW geniculate wing VPL lateral part of the ventroposteriornucleus

VPM medial part of the ventroposteriornucleus

278 Cat Cortico-thalamic Network • Scannell et al.

1995). In some cases, this is due to methodological differences.

Connection tracing (the bulk of the data used in this study), cyto- or

chemo-architectonics and electrophysiological mapping do not always

yield divisions that map neatly onto one another. These problems made

collation particularly difficult for connections of the anterior parts of

the lateral posterior complex, the posterior group of nuclei and the

suprageniculate region. Difficulties were also encountered in the lateral

posterior complex, the medial geniculate complex, the posterior group

and the ventroposterior complex, where there is controversy or

ambiguity as to the ‘best’ parcellation.

Collation of Connection Data

We produced a database of thalamo-cortico-cortical connections from

information in a large number of published studies of connectivity in the

adult cat. Only data from the cat were used and no extrapolations were

made from structures thought to be homologous in other species. We

considered data in the text and figures of papers on cat cortico-cortical,

cortico-thalamic and thalamo-cortical connectivity. For each reported

instance of a connection, the database gives the origin and target regions,

the weight or strength (see below), local origin or termination patterns

(where known), and the study from which the data were taken. For some

connections, the database has notes on alternative parcellation schemes.

Connections that were reported as dense or strong were given a strength

weighting of 3; weak or sparse connections were weighted as 1; and

connections of intermediate strength, or for which no strength

information was available, were weighted as 2. We assumed that

descriptions of the density of projections as strong, moderate or weak

made in different studies were equivalent.

Translation from Database to Connection Matrix

The areas that we have chosen to include in our connection matrix ref lect

current connection data. We have balanced the need to avoid ‘lumping’

together structures that are known to be distinct, with the need to have

sufficient data for each structure listed in the connection matrix. Where

the database contains fine areal or nuclear divisions that have not yet been

resolved in a substantial number of connection tracing studies, we have

tended to ‘lump’ them together for the analyses. For example, Clasca et al.

(1997) present compelling evidence that our cortical areas Ia and Ig both

contain a number of distinct cortical fields. However, there are no

published cortical connection data for these divisions at present, so we

include Clasca et al.’s areas ALd, ALv and PI within our area Ia, and Clasca

et al.’s areas AS, DI and GI within our area Ig. We hope that our

parcellation will improve as new data become available.

The database was used to produce our best estimate of thalamo-

cortical connectivity, shown in Table 2. Where there was any doubt in

translation between the database and connection matrix, we checked the

original papers and reconsidered both text and figures. The translation

process was difficult when the reports in the literature could not be

unambiguously assigned to a single one of our areas or nuclei. In these

cases, we were careful to balance the need to include as many reported

connections as possible, with the need to avoid ‘extrapolating’

connections to areas or nuclei which do not, in fact, possess them. For

example, if one study identified a connection from a cortical area to the

‘lateral posterior nucleus’, LP, while other studies identified connections

to specific nuclei within LP (e.g. LPl and LPm), then we would to assign

the connections to LPl and LPm only.

Occasionally, given all the information in the database, a connection

could still not be assigned unambiguously to one of our areas or nuclei. In

these cases, where reasonable, we would assign the connection to our

areas included within the larger, ambiguous region. For example, if the

only available information showed that Sg projected to areas 35 and/or

36, then we would assign the connection to both 35 and 36. We are aware

that this risks one false positive connection. However, the alternative

(ignoring Sg’s projection to 35 and/or 36) is certain to incur at least one,

and may incur two, false negatives. In practice, we applied this reasoning

to a relatively small number of connections, which disproportionately

affected areas 35 and 36, and nuclei PARA and PARP. Areas 35 and 36 are

thin, adjacent strips of cortex that are difficult to inject independently.

PARA and PARP are small, adjacent thalamic nuclei that are difficult to

distinguish in connection tracing studies. Therefore, there is likely to be

some ‘cross-contamination’ in our connection matrix and analyses,

between areas 35 and 36, and between thalamic nuclei PARA and PARP.

All measurements in all areas of science are likely to contain a degree

of error. This is true for our database and our translation from database to

connection matrix. While we have make great efforts to be as accurate as

possible, we encourage readers to check the primary literature before

uncritically accepting connections presented here.

Missing Connections

In principle, our analyses could distinguish between connections that

have been explicitly reported as absent and connections for which there

are no explicit reports of presence or absence. For several reasons, we

have chosen to analyse ‘confirmed absent’ and ‘unreported’ connections

as if they were equivalent. First, most areas or nuclei only connect with a

small fraction, 5–25%, of the other structures, so non-connections are the

norm. Non-connections are rarely reported explicitly unless they are

unexpected, controversial or otherwise of particular interest. Therefore,

there are few explicitly confirmed absent connections in the cat.

However, a substantial proportion of studies in the cat inspect large

regions of the brain (sometimes the entire cortex or thalamus) for labelled

cells or terminals. We take the lack of reports of connections in the text

or figures of such papers as implicit evidence of absence and think it

unlikely that there are many substantial undiscovered connections

between well-studied regions of the brain. Of course, this may change

with the introduction more sensitive connection tracing methods, or

more systematic attention to very small quantities of transported tracer

that may currently be dismissed as ‘background’ labelling.

Second, many ‘missing’ connections belong to poorly studied regions

of the brain. Therefore, we have chosen to omit some of the less studied

areas from the analysis. We also think that ‘missing’ connections are most

likely to be projections to regions of the brain that have received few

tracer injections. This is because retrograde tracing studies are more

common than anterograde tracing studies and projections from the

poorly studied brain regions show up when the better studied regions are

injected.

Third, it is likely that thalamo-cortical connections show complete

reciprocity (e.g. Raczkowski and Rosenquist, 1983; Burton and Kopf,

1984b; Symonds and Rosenquist, 1984; Jones, 1985). Therefore, assu-

ming thalamo-cortical reciprocity should give a more accurate picture

than using only connections that have been directly demonstrated. For

example, many thalamo-cortical connections, found with cortical injec-

tions of retrograde tracers, have not been studied using cortical injections

of anterograde tracers, or thalamic injection of retrograde tracers. To

avoid ignoring such connections, our analyses assume that all con-

nections between cortex and thalamus are strictly reciprocal.

Fourth, to see if ‘missing’ cortical reciprocal connections in Table 2

have a major effect on the results of our analyses, we have produced, and

analysed with non-metric multi-dimensional scaling (NMDS), alternative

matrices that assume cortical connections are also reciprocal. First, a

symmetrized matrix, ‘sym’, was produced by ref lecting the ‘raw’ matrix

(Table 2) about its leading diagonal and then adding the ref lected matrix

to the original. Second, a ‘winner takes all’ matrix, ‘wta’, was produced by

ref lecting the raw matrix about its leading diagonal and then adding it

back to the original in a ‘winner takes all’ fashion, so that the strongest

connection at any site was retained. The raw matrix (Table 1) and the wta

matrix have four ranks of connection strength: 0 (absent or unreported),

1 (weak or sparse), 2 (moderate) and 3 (strong or dense). The sym matrix

has seven ranks of connection strength (0–6).

Fifth, Young (1992) and Young et al. (1995) have examined

quantitatively the effects of different ways of treating non-connections

and found that they have a relatively small effect on the overall solution.

Few connections in the cat have been explicitly reported as absent, so

adding an extra rank to the connection matrix (Table 2) to distinguish the

few ‘confirmed absent’ connections from the many ‘presumed absent’

connections would have very little effect. We also note that all our

analyses on n structures seek to satisfy (n2 – n) connectional constraints

while each structure contributes only (2n – 1) connectional constraints.

Therefore, major changes in the connection patterns of a few poorly

studied structures will have a major impact on those structures in any

connectional analysis, but will have very little impact on the overall

analysis (Scannell et al., 1995; Young et al., 1995).

Cerebral Cortex Apr/May 1999, V 9 N 3 279

We think that this approach is reasonable and we prefer it to

mathematically based missing data estimation methods (Jouve et al.,

1998). The issue of missing connections can be properly resolved only by

further anatomical study.

Analysis Methods

Large connection matrices are too complicated for analysis with unaided

intuition. We have used three methods to summarize and represent the

connection data in a more transparent way. The first, and most widely

used to date (e.g. Young, 1992, 1993; Scannell and Young, 1993; Scannell

et al., 1995; Young et al., 1995a,b,c), is NMDS. Simmen et al. (1994) and

Goodhill et al. (1995) have criticized the use of NMDS on anatomical

connection data. While we think that NMDS can recover structure from

anatomical connection data (Young et al., 1994, 1995c), the comments of

Simmen et al. (1994) and Goodhill et al. (1995) have been very useful.

They have encouraged us to develop and apply a wide range of analysis

methods to help distinguish real connectional features from the

peculiarities of individual analysis methods. Therefore, we have analysed

connection data with NMDS, with non-parametric cluster analysis and

with a novel method, optimal set analysis.

Non-metric Multidimensional Scaling

NMDS is a method that finds a configuration of points (in this case

cortical areas and thalamic nuclei) that matches a set of experimentally

measured ‘proximities’ (in this case anatomical connections). In a

theoretically perfect arrangement, all areas or nuclei that are connected

by a connection weighted ‘3’ should be closer to each other than all areas

linked by a connection weighted ‘2’, which should in turn be closer than

areas linked by a ‘1’, and so on. Such an ideal configuration would

perfectly capture the connectional proximities present in the connection

data — a process exactly analogous to deriving a map from a mileage chart

in a road atlas.

NMDS is a form of data compression in which a large set of ordinal

measures (e.g. 95 × 95 connection matrix) is represented as a smaller set

of metric measures (e.g. 95 × 3 set of coordinates). This kind of data

compression is not without problems. For example, if the proximity

matrix is asymmetric (e.g. Table 2), the connectional constraints cannot

be perfectly satisfied by a spatial arrangement of points. Nevertheless,

extensive simulation studies (Young et al., 1994, 1995c), have shown that

this approach performs excellently in recovering metric structure from

test data that matches as closely as possible the structure of neuro-

anatomical connection data. Analyses of test data designed to match the

cat thalamo-cortical connection data has found that cat thalamo-cortical

connection data behaves like ‘noisy’ test data generated in between four

and six dimensions (Scannell et al., 1997; J. W. Scannell, unpublished

observations). NMDS configurations derived from such test data in

the appropriate number of dimensions (i.e. five dimensions for five-

dimensional test data) typically account for ∼90% of the variability of the

original test structure. For practical purposes, we are limited to

presenting three-dimensional configurations, which typically account

for ∼50% of the variability of appropriate five-dimensional test data

(J.W. Scannell, unpublished observations).

Young et al. (1995c) have shown that for appropriate test data, some

matrix transforms can slightly enhance the ability of NMDS to recover

metric structure. Consequently, we have performed the wdsm1 trans-

form (Young et al., 1995c) prior to NMDS for all the cortical systems

illustrated in this paper. The high degree of congruence between

solutions of the wdsm1-transformed matrices and the solutions derived

from the other matrices is shown in Table 3, and is illustrated in Figures 1

and 2 which show configurations derived with ALSCAL from wdsm1,

raw and sym matrices. Configurations derived with the other NMDS

approaches and have been inspected using computer animation.

Configurations of points derived from low-level ordinal data vary with

NMDS method. Previous studies (e.g. Young, 1992, 1993; Scannell and

Young, 1993) have derived two-dimensional configurations with the

ALSCAL algorithm (Takane et al., 1977; Young et al., 1978) in the SPSS

statistical environment (Young and Harris, 1990). In this paper we

present stereograms of three-dimensional configurations as they capture

more of the data structure. The configurations were derived using a

variety of NMDS methods to see which features of the configurations are

robust and consistent properties of the connection data. The methods

were ‘MDS’ in the SAS statistical programming environment and ‘ALSCAL’

in the SPSS environment. All connection data (raw, sym, wta and wdsm1)

were scaled at the ordinal level of measurement using the secondary

approach to ties (Kruskal, 1964a,b; Young et al., 1995c). SAS solutions

were produced using options ‘FIT = 1’ (fits the data to distances) and ‘FIT

= 2’ (fits the data to squared distances). ‘FIT = 2’ emphasizes the

importance of long distances (i.e. non-connections) and produces output

similar to ALSCAL.

NMDS solutions can become trapped in local minima. To guard

against this, we shuff led the order of areas in each matrix while keeping

the connections the same. Shuff ling input order had no effect on all

solutions produced with MDS in SAS, and on solutions derived from

symmetric matrices (sym, wta, wdsm1) with ALSCAL in SPSS. However,

shuff ling input order did inf luence the apparent fit of solutions derived

from raw matrices with ALSCAL in SPSS. In these cases, we found that a

‘seeding’ method was effective in finding low-dimensional configurations

with relatively high apparent fit. Configurations derived from wta or sym

matrices were used to provide the initial coordinates for scaling the raw

matrix. The resulting raw configurations reached convergence in fewer

iterations and had higher apparent fit than configurations derived from

random co-ordinates.

NMDS configurations can be compared quantitatively using a

regression-like procedure, Procrustes’ rotation (Schonemann and Carrol,

1970; Gower, 1971). Procrustes’ rotation, implemented in the GENSTAT

statistical programming language, can compare structures in any number

of dimensions by rigidly ref lecting, rotating and then scaling them to find

an optimal fit. After finding the best rigid transform, Procrustes yields a

variance explained statistic (R2) that ref lects the goodness of fit. The

statistical rarity of each comparison may be evaluated using a

randomization test (Edgington, 1980) in which one of the structures is

repeatedly randomly shuff led and the R2 statistic recalculated.

Optimal Set Analysis

Optimal set analysis is an alternative and independent method for

analysing global connectional architecture. This approach views a ‘neural

system’ as something containing components more connected with each

other than with the components of other systems. So, for example, the

visual system might be thought of as the collection of cortical areas and

thalamic nuclei that are highly interconnected with each other and

relatively weakly connected with the other cortical systems. We wanted

to identify the systems within the thalamo-cortical network by arranging

areas and nuclei into sets with as few as possible connections between

the sets and as few as possible non-connections within the sets. We

wanted to let the connection data itself decide the number and the

composition of the sets of areas and nuclei.

Optimal set analysis presents a formidable computational problem so

connection data were analysed using CANTOR, a custom-written

software system for statistical optimization. The CANTOR system consists

of a database of objects, linked to each other through relational functions.

In the present application, the CANTOR objects represented cortical

areas and thalamic nuclei, while the relational functions represented their

anatomical interconnections. CANTOR arranges the objects by stochastic

optimisation to minimize the cost of the relational functions in a way that

is chosen by the experimenter. We chose costs that matched our idea of a

neural system; a thing containing components more connected with each

other than with the components of other systems. Therefore, the cost

measured non-connections within sets and connections between sets.

The CANTOR environment provides routines that permit the

rearrangement of the objects within the database to minimize costs by

evolutionary optimization, a method based on simulated annealing (Van

Laarhoven and Aarts, 1987). More technical details of the optimization

algorithm adapted for connectivity analyses are described by Hilgetag et

al. (1996, 1997b, 1998). The software is written in ANSI-C and has been

implemented on a wide variety of Unix computer systems. Special

memory management ensures that the large quantities of complex data

can be analysed on medium-sized workstations. Our computations are

run on a DEC Alpha 3000–600 workstation with 256 MB of RAM. The

runs typically require several hours to weeks of CPU time, depending on

the size of the area-to-area connectivity matrices and the thoroughness of

search for optimal solutions.

In this study, CANTOR minimized a two-component integer cost

280 Cat Cortico-thalamic Network • Scannell et al.

function to find optimal connectional sets within the thalamo-cortical

system. The cost consisted two terms. The first was a repulsion term that

counted non-existing connections within the potential sets. Optimizing

the repulsion component alone breaks larger clusters into smaller

clusters. The second cost consisted of an attraction term counting

existing connections between sets. Optimizing the attraction term alone

will, in the limit case, result in one large cluster containing all the areas

and nuclei.

The costs of existing connections between sets were proportional to

the connection weights (‘1’, ‘2’ or ‘3’). This cost function assumes an

arithmetic relationship between connection ranks, in contrast to NMDS

which simply assumes a monotonic relationship between connection

ranks. Starting from a random assignment of areas to sets, the number and

structure of the sets were changed iteratively through relocation of one

area to another set or by swapping of two areas belonging to different

sets. Thresholds were set to ensure that structures made stepwise

improvements. These set manipulations and threshold parameters

ensured that the space of candidate arrangements was searched at the

smallest possible resolution and that the search did not become trapped

in local minima. Hence, the CANTOR system determined automatically

the number and composition of sets for the optimal overall arrangement

of the areas and nuclei of the thalamo-cortical system.

We changed the number and composition of the optimal sets by

varying the strength of attraction and repulsion in the cost function.

Repulsion ranged from 30 to 1 (cost of non-connections within clusters),

with attraction set to 1. Attraction ranged from 1 to 10 (the attraction

term is multiplied by connection weight to compute the cost of

connections between clusters), while repulsion was held at 1. Increasing

the attraction yields larger sets, or only one in the limit case of

overwhelming attraction. Reducing attraction yields more, smaller sets.

These manipulations make it possible to explore the cluster structure of

the thalamo-cortical network over a range of scales. They emphasize

different aspects of the organization of the network. Perhaps the most

intuitively obvious, and the one shown in Figures 3 and 4 is the case

where attraction and repulsion are both set to 1. Here, the cost of

individual connections between clusters equals their connection weight

and the cost of non-connections within clusters equals 1. Since CANTOR

typically finds many different solutions with the same or similar low cost,

the results are summarized in a matrix showing the proportion of the

solutions in which two structures occupied the same set.

We estimated the statistical significance of the results of optimal set

analysis by comparing the cost of solutions calculated from connection

data with the cost of equivalent solutions calculated from randomized

connection matrices. The cost of solutions derived from connection data

was comparable with the cost of solutions calculated from four- to

six-dimensional test data (Scannell et al., 1997; J.W. Scannell, unpublished

observations).

Non-parametric Cluster Analysis

Cluster analysis is a statistical method that finds groups within a data set.

It has already been used in the analysis of connection data by Musil and

Olson (1991) to delineate the areas in the medial frontal cortex of the cat,

and by Sherk (1986) to map the suprasylvian region of the cat. Here, the

rationale behind our use of cluster analysis was similar to that behind

optimal set analysis, i.e. to find groups of connectionally associated areas

and nuclei (see Hilgetag et al., 1998).

To see the robust and reliable groupings within the thalamo-cortical

network, we performed a variety of cluster analyses on a variety of NMDS

configurations. We used the SAS/STAT system to make several five-

dimensional NMDS configurations from the raw connection matrix,

setting the FIT option to 0.5, 1 and 2, with both the primary and

secondary approach to ties. NMDS analyses in five dimensions were

chosen for two reasons. First, the accuracy of density estimation falls with

increased dimensionality (Epanechnikov, 1969, Silvermann, 1986).

Second, studies with test data indicate that NMDS in five dimensions is

adequate to reasonably represent the cat thalamo-cortical connection data

(J.W. Scannell, unpublished observations). We then analysed the config-

urations using the MODECLUS non-parametric cluster analysis function in

the SAS/STAT system.

Before clustering, MODECLUS calculates the probability density

within the NMDS configuration using either fixed or variable radius

spherical kernels. With fixed radius kernels, all points have identical

kernel radii. The variable radius kernel approach is based on finding the

distance to the ‘kth nearest neighbour’. Once a density function has been

estimated for a set of points (in this case, representing cortical areas and

thalamic nuclei), MODECLUS detects the local maxima of the density

function to define the cluster structure of the data. When fixed radii

kernels are used, MODECLUS can then test the significance of clusters by

comparing the maximum estimated density at points within the cluster to

the maximum estimated density at the cluster’s boundary.

Since cluster analysis is adept at finding clusters in data where no real

clusters exist (Gordon, 1996), we ran ten kinds of MODECLUS analysis

paradigm on each of the six NMDS configurations and then pooled the

results. We did this to ‘average out’ the spurious clusters. We also ensured

that our cluster paradigms used a wide range of kernel radii so that both

local and large-scale cluster structure were explored. The first paradigm

ran 30 cluster analyses with increasing fixed-radius density estimation

and clustering kernels. The radius of this kernel ranged from the

minimum interpoint distance in the configuration to the mean interpoint

distance in the configuration. The second, third and fourth paradigms

used a fixed radius kernel that was calculated from the cluster analysis

under the first paradigm. The kernel radius was chosen to produce an

initial number of clusters equal to the total number of points in the

analysis divided by 2, 4 and 8 respectively. A hierarchical cluster tree was

obtained for each scheme by testing the significance of the clusters. This

procedure compared the maximum estimated density of points within a

cluster to the maximum around the cluster’s border (referred to here as

the ‘saddle point’) in order to estimate the cluster’s significance. On

successive iterations, the least significant cluster was dissolved or their

members assigned to a neighbouring cluster if it was close enough. The

fifth, sixth and seventh paradigms were based on nearest-neighbour

kernels with K values set to 2, 3 and 4 (i.e. the density estimation and

clustering kernels were set to the value so that the number of neighbours

of each point was a minimum of 1, 2 and 3). The eight, ninth and tenth

paradigms used a nearest-neighbour clustering method and gave a

hierarchically organized cluster scheme. This involved the use of

fixed-radius density estimation kernel (with radius set to the value of the

configuration’s mean distance) and a clustering kernel based on a

nearest-neighbour CK value set to 2, 3 and 4.

Since the series of cluster analyses finds many different solutions, the

results are summarized in a matrix showing the proportion of the solu-

tions in which two structures occupied the same cluster. The summary

matrix (Figure 6) was corrected so that each paradigm contributed

equally to the cluster-count score. Note that a structure from a cluster that

had been dissolved (e.g. paradigms 2, 3 and 4) would not be counted as

being in the same cluster as any other structures, including itself. Thus

there may be cluster counts along the leading diagonal of Figure 4 of

<100%.

Results

Parcellation and Connections

Table 2 summarizes the literature on the extrinsic connectivity

of the thalamo-cortical network. It is drawn from the full

database (see http://www.psychology.ncl.ac.uk/jack/cor_thal.

html), which provides considerably more detailed information.

The connection strengths are on an ordinal scale that assumes

only that the relationship between the ranks is monotonic; a ‘3’

is stronger than a ‘2’ which is stronger than a ‘1’ which is

stronger than a ‘0’. In quantitative terms, connections weighted

‘3’ may be 2 or 3 orders of magnitude stronger than connections

weighted ‘1’ (see e.g. Musil and Olson, 1988a,b). Table 2 shows

826 cortico-cortical connections. Each cortical area connects

with, on average, 15 other areas (SD = 8). This level of

connectivity corresponds to 28% of possible cortico-cortical

connections. Table 2 also shows 651 connections between

thalamus and cortex, which we assume are wholly reciprocal

(Raczkowski and Rosenquist, 1983; Burton and Kopf, 1984b;

Symonds and Rosenquist, 1984; Jones, 1985). On average, each

Cerebral Cortex Apr/May 1999, V 9 N 3 281

282 Cat Cortico-thalamic Network • Scannell et al.

thalamic nucleus projects to, and receives projections from, 15

cortical areas (SD = 10). This constitutes 28% of all possible

thalamo-cortical connections, were each nucleus to project to

each cortical area. There are no direct connections between the

thalamic nuclei.

The connection density summary of Table 2 shows that there

is great variability in the numbers of connections and the

distribution of connection weights between areas and nuclei.

Some structures (e.g. 17, SII) have a small number of dense

projections, while others (e.g. CMN, CLN, CGp and AES) have

much more widespread connectivity. In general, however, weak

connections are much more common than strong connections.

While some cortical areas and thalamic nuclei have received

vast amounts of experimental attention (e.g. 17, LG, PMLS),

some have been the focus of much less research. Further work is

required on parcellation in the posterior complex (POi, POl,

POm), the suprageniculate region (Sg, LM-Sg) and the anterior

parts of the lateral-posterior complex (LIc, LIo). Data are sparse

for cortical connections to the midline and intralaminar nuclei,

connections to and from the anterior group of nuclei, and con-

nections to and from the thalamic regions where parcellation is

unclear. Therefore, we advise caution in interpreting the

database for these structures. Structures with very sparsely

documented connectivity can be problematic for our quanti-

tative analyses. Consequently, POm, POl, POi, LIc and LIo

occupy peripheral or erratic locations in some of the figures.

NMDS Analyses

NMDS provided reasonable three-dimensional representations of

the connection data as the apparent fit of the configurations

derived from connection data was much better than the fit from

appropriate random test data (typically by 10 or more z-scores,

P < 10–12). The apparent fit of configurations derived from

connection data were comparable with the fit derived from

appropriate four- to six-dimensional test data (Scannell et al.,

1997; J.W. Scannell, unpublished observations).

Congruence of Different NMDS Approaches

Table 3 shows that across all input matrices the different NMDS

methods yield quantitatively similar structures (R2 varies

between 0.78 and 1.00). We have used computer animation to

visualize all the structures and the similarity of the qualitative

impression is even greater than the quantitative similarity might

suggest. In general, MDS with ‘FIT = 2’ and ALSCAL, when using

low-level ordinal data (wta, sym and raw), give configurations in

which the areas and nuclei lie in a shell around a hollow region.

This is because these scaling methods emphasize the non-

connections, or long distances (Takane et al., 1977; Young et al.,

1978), and the arrangement of points that has the most long

inter-point distances is a hollow shell (Fig. 1). In contrast, MDS

with ‘FIT = 1’ (Fig. 2) and all analyses with the wdsm1 matrix

(e.g. Fig. 1) give configurations with points more evenly

distributed within a volume of space. MDS with ‘FIT = 1’ fits the

Table 3Different NMDS methods yield congruent result

Algorithm

ALSCAL ALSCAL ALSCAL MDS MDS MDS MDS MDS MDS MDS MDS MDS MDS MDS MDS

Matrix

raw sym wta wdsm1sym

wdsm1sym

wdsm1wta

wdsm1wta

wdsm1raw

wdsm1raw

raw raw sym sym wta wta

Fit

Algorithm Matrix Fit 2 1 2 1 2 1 2 1 2 1 2 1

ALSCAL raw 1.00 * * * * * * * * * * * * * *ALSCAL sym 1.00 1.00 * * * * * * * * * * * * *ALSCAL wta 0.99 1.00 1.00 * * * * * * * * * * * *MDS wdsm1 sym 2 0.88 0.89 0.89 1.00 * * * * * * * * * * *MDS wdsm1 sym 1 0.84 0.85 0.85 0.98 1.00 * * * * * * * *MDS wdsm1 wta 2 0.86 0.87 0.88 1.00 0.97 1.00 * * * * * * * * *MDS wdsrn1 wta 1 0.83 0.84 0.84 0.97 1.00 0.97 1.00 * * * * * * * *MDS wdsm1 raw 2 0.88 0.89 0.89 1.00 0.98 1.00 0.97 1.00 * * * * * * *MDS wdsm1 raw 1 0.84 0.85 0.85 0.98 1.00 0.97 1.00 0.98 1.00 * * * * * *MDS raw 2 0.99 0.99 0.99 0.90 0.86 0.88 0.85 0.90 0.86 1.00 * * * * *MDS raw 1 0.88 0.88 0.87 0.79 0.79 0.78 0.78 0.79 0.79 0.88 1.00 * * * *MDS sym 2 0.99 0.99 0.99 0.90 0.87 0.89 0.86 0.90 0.87 1.00 0.88 1.00 * * *MDS sym 1 0.88 0.88 0.88 0.80 0.81 0.80 0.80 0.80 0.81 0.88 0.97 0.89 1.00 * *MDS wta 2 0.98 0.98 0.99 0.91 0.87 0.90 0.86 0.91 0.87 0.99 0.87 1.00 0.88 1.00MDS wta 1 0.86 0.86 0.87 0.80 0.80 0.80 0.80 0.80 0.80 0.87 0.94 0.87 0.97 0.88 1.00

The table shows the goodness of fit (variance explained, R2, from Procrustes’ rotation) between configurations derived in three dimensions using a variety of NMDS methods on the raw, symmetrized,winner-takes-all and wdsm1 transformed connection matrices (‘raw’, ‘sym’, ‘wta’ and ‘wdsm1’). ‘ALSCAL’ and ‘MDS’ show whether configurations were derived using ALSCAL or MDS. ‘f1’ and ‘f2’ werederived using ‘FIT=1’ and ‘FIT=2’ respectively. The table shows that all NMDS methods yield quantitatively similar results.

←

Table 2Matrix of cortico-cortical, thalamo-cortical and cortico-thalamic connections in the cat. Connections weighted ‘3’ are strong, connections weighted ‘2’ are intermediate, connections weighted ‘1’ are weakor sparse, and connections marked ‘.’ have been explicitly reported as absent or are unreported. The upper square of the table shows cortico-cortical connections. Reading horizontally in this part of thetable, from a cortical structure, shows afferent connections and reading vertically shows efferent connections. The lower rectangle of the table shows cortico-thalamic and thalamo-cortical connections,which we assume are invariably reciprocal (Raczkowski and Rosenquist, 1983; Burton and Kopf, 1984b; Symonds and Rosenquist, 1984; Jones, 1985). The right margin of the table summarizes thefrequency distribution of connection weighted 3, 2 and 1 for each anatomical structure. Afferent and efferent connections have been pooled. Darker shades indicate a greater number of connections. Thearea to area connection data in this table formed the sole input to the NMDS, optimal set, and non-parametric cluster analyses.

Cerebral Cortex Apr/May 1999, V 9 N 3 283

experimental disparities (connection data) to distance, so does

not emphasize non-connections. The wdsm1 transform inter-

polates extra ordinal ranks, converting the low-level connection

matrix into a higher-level ‘similarity of connection’ matrix. This

removes the large number of identical long distances (non-

connections) that are present in the raw data.

Figure 1 shows the two least similar NMDS configurations

(wdsm1 wta with ‘FIT=2’ and MDS raw with ‘FIT=1’, which

account for 78% of each other’s variability by Procrustes’

rotation). Despite the fact that they are the most quantitatively

dissimilar configurations, they give similar qualitative impres-

sions of the connectional architecture of the thalamo-cortical

network.

The qualitative similarity of the NMDS configurations (Table

3, Figs 1, 2 and 4) arises because the local relations between the

points are relatively conserved despite the fact that the

configurations occupy different shaped volumes. For example,

Figures 1 and 2 show that the hippocampus is always distant

from the sensory periphery, that the four major cortical systems

are distinguishable, and that the visual system shows some

divergence at higher levels into structures associated with the

somato-motor system (AES, area 7) and structures associated

with the fronto-limbic system (20b, PS). Some of the midline and

intralaminar nuclei (CLN, CMN, PAC, PF, CM) do not appear to

be closely allied with any particular system, but project widely to

a large number of cortical areas. The remaining midline and

intralaminar nuclei (RH, MV-RE, PAT, PARA, PARP) appear to be

associated with the fronto-limbic complex.

Figure 1. Connectional relationships in the thalamo-cortical network. The figure is a stereogram showing the two least similar three-dimensional configurations yielded by the varietyof NMDS analyses of the different versions of the connection matrix. The first configuration (top) was derived using MDS on the wta wdsm1 matrix with FIT=2, and the second(bottom) was derived using MDS on the raw matrix with FIT=1. The R2, by Procrustes’ rotation, between the configurations is 0.778. Anatomical structures within the visual systemare coloured blue, auditory structures are green, somato-motor structures are red and fronto-limbic structures are brown. CLN, CMN, PAC, PF and CM are black. The figure wascomputed using all the connection data but, for clarity, only strong reciprocal connections, with a sum of weights in both directions >4, are drawn. All structures have someconnections (see Table 2). The different NMDS approaches arrange the points in different volumes of space, but the local organization within systems and the relationships betweenthe systems are broadly similar. The apparent badness of fit (Young et al. 1995a) for scaling the wdsm1 matrix was 0. 21, with a distance correlation of 0.94. The apparent badnessof fit for scaling the raw matrix was 0.26 with a distance correlation of 0.61. The stereograms in this paper may be free-fused to give three-dimensional images. Divergent viewingmay be easier for those used to ‘magic eye’ pictures. To view, hold the figure at arm’s length in front of the face, just below the line of sight. Look steadily at a distant object, somethingoutside the window, that is just visible over the top of the page, then switch your view to the page without converging the eyes. It helps if the hands are kept clear of the figures whileattempting to fuse them. One may practice by drawing two dots with a felt-tipped pen ∼4 cm apart at the very top of piece of paper. Repeat viewing the procedure above. At first,you should see four dots at the top of the paper just below the distant object that you are fixating. If you switch your view from the distant object to the dots, the central two dotsshould move towards each other and then merge when you have achieved divergent fixation. Practice with dots that are further apart until it is possible to fuse the figure. For practice,see Johnstone (1995).

284 Cat Cortico-thalamic Network • Scannell et al.

NMDS Representations of the Thalamo-cortical Network

The next section outlines NMDS analyses of the thalamo-cortical

network and focuses on the visual, auditory, somato-motor,

fronto-limbic and global systems, and on the midline and

intralaminar nuclei. We report only those characteristics of the

connectional architecture that appear robust and independent

of the details of the NMDS method. The NMDS configurations

here can best be regarded as provisional attempts at a systematic

map of the thalamo-cortical network. Like any early map, they

may provide a useful guide, but will also contain omissions and

distortions that may be improved with better analysis methods

and anatomical data.

Global Thalamo-cortico-cortical System

Figures 1 and 2 show that the basic connectional plan found in

the cortex, of three broadly hierarchical sensory/sensory-motor

systems and a fourth fronto-limbic system (Scannell and Young,

1993; Young, 1993), also provides a good summary description

of the thalamo-cortical network. The wdsm1 configuration in

Figure 1, for example, arranges points representing cortical

areas, and thalamic nuclei fall in a region of space that is

roughly the shape of a triangular pyramid. Three of the corners

correspond to the most topologically peripheral parts of the

sensory and sensory/motor systems, while the fourth corner

corresponds to the most topologically central parts of the fronto-

limbic complex; the hippocampus and associated structures.

The central region of the pyramid (Fig. 1) where the four systems

meet contains, as might be expected, those regions of cortex

where information from the different sensory modalities

converges (e.g. EPp, 7, AES, Ia, Ig, CGp; 36).

Figure 2 shows all the documented connections of the

thalamo-cortical network and gives some indication of the

complexity of the thalamo-cortical network even at the crude

level of organization presented here. The complexity of Figure 2

shows why systematic collation and data analysis may prove

useful tools in the interpretation of primary neuroanatomical

data.

Visual System

Figures 1 and 2 illustrate the connectional architecture of the cat

Figure 2. Connectional relationships in the thalamo-cortical network. The figure shows the three-dimensional NMDS configuration derived using MDS with ‘FIT=2’ on the rawmatrix. Anatomical structures within the visual system are coloured blue, auditory structures are green, somato-motor structures are red and fronto-limbic structures are brown. CLN,CMN, PAC, PF and CM are black. The figure was computed using all the connection data. The top figure illustrates only those connections where the sum of the weights in bothdirections is ≥4. The lower figure shows all connections, and graphically illustrates the great complexity of the network of extrinsic connections between cortical areas and thalamicnuclei. Stronger connections are darker grey and weaker connections are lighter grey. The apparent badness of fit of the configuration (Young et al., 1995) is 0.41 and the distancecorrelation 0.61.

Cerebral Cortex Apr/May 1999, V 9 N 3 285

thalamo-cortical visual system (blue). Figures 1 and 2 show

that the visual system has a complex, yet broadly hierarchical

organization. The periphery of the thalamo-cortical visual system

are the components of the lateral geniculate nucleus which are

in close connectional association with areas 17 and 18. Higher

are areas 19, 21a and 21b, which are connectionally close to LPl.

Higher still, there appears to be some divergence in the visual

system; PLLS and ALLS lead towards AES, 7 and 5Bl, which are in

close connectional association with the somato-motor system.

LM-Sg is closely associated with AES and Ig. In contrast, PS, 20a

and 20b lead towards 35 and 36, RS and limbic structures. Some

of these features are not apparent in Figure 2 from this particular

view, but are clearer at when the configuration is viewed from

other directions.

The connectional divergence in higher visual areas is

ref lected in a functional division into areas concerned with

motion and visuo-spatial functions (e.g. 7, AES) and those that

are involved in static object vision (e.g. 20a, 20b) (Lomber et al.,

1996a,b; Scannell et al., 1996). This arrangement is analogous

to the functional (Ungerleider and Mishkin, 1982) and

connectional (Young, 1992) divergence into somewhat distinct

object versus motion and visuo-spatial streams found in the

macaque visual system. Cats and macaques diverged from a small

insectivorous ancestor over 60 million years ago (Savage, 1986),

so the similarities in visual system organization may well ref lect

convergent evolution rather than homology. This, in turn, could

ref lect a wiring efficiency advantage in separating form and

motion vision (Jacobs and Jordan, 1992).

Auditory System

Figures 1 and 2 illustrate the connectional architecture of

the thalamo-cortical auditory system (green). The auditory

structures occupy a tight cluster, so the system appears less

hierarchical than the visual system and somato-motor systems.

MGV, MGDdd and MGDd constitute the periphery of the

thalamo-cortical auditory system. These structures are in close

connectional association with the tonotopically organized ‘core’

fields VPc, AAF and AI (Brugge and Reale, 1985; Morel and Imig,

1987). Associated with the higher auditory area, Tem, is MGvl.

The highest stations in the auditory system appear to include

POl, SG and EPp, which are in close association with Ia, Ig, 35

and 36. MGDds tends to occupy a rather peripheral position,

which may be an artefactual consequence of the fact that it has

few documented connections.

Somato-motor System

Figures 1 and 2 show that area-to-area connection patterns define

a single somato-motor network (red) that contains both sensory

and motor structures in close connectional association. In terms

of extrinsic thalamo-cortical connection patterns, motor and

somatosensory structures form a single unit. The somato-motor

system is complex yet relatively hierarchical. VPL and VPM,

which are closely associated with the primary somatosensory

areas 1, 2, 3b and 3a, form the periphery of the somato-motor

network. These areas are very closely associated with the

primary motor cortex, area SII and some parts of area 5. VA-VL

is ‘higher’ than VPL and VPM and is closely associated with

areas 4 and 6. VMP is poised between the somato-motor and

fronto-limbic systems. Higher somato-motor areas, in particular

6l and 6m, are very closely associated with fronto-limbic

structures such as CGa, CGp, Ia and Ig. 5Bl, the most lateral part

of posterior area 5, is closely associated with visuo-motor

structures such as AES and 7, suggesting a possible visual role for

this region.

Fronto-limbic System

We use the term ‘fronto-limbic system’ to describe a result that

we obtain when we analyse connection data with several formal

methods, namely the tendency of areas of the prefrontal cortex,

‘limbic system’ and associated thalamic nuclei to group together

(Scannell and Young, 1993; Young, 1993). Systematic and

principled analysis of connection data decides that these

structures belong together. We simply report the association and

give it a name. We are aware that the term ‘limbic system’ is

problematic (e.g. Kotter, 1992) so our use of the term

‘fronto-limbic’ does not imply any particular functional inter-

pretation. Fronto-limbic structures include cingulate, insular,

prefrontal, perirhinal, entorhinal and retrosplenial cortex,

the subiculum, presubiculum, hippocampus, amygdala and

associated thalamic nuclei such as MD.

From Figures 1 and 2 it is clear that areas 35, 36, Ia, Ig, CGp

and CGa are connectionally interposed between the cortical

sensory and sensory-motor systems and the ‘deeper’ parts of the

fronto-limbic complex. Most distant from the sensory periphery

are ER, SB and Hipp. As would be expected, MD is in close

association with prefrontal cortical areas, and MV-RE with the

components of the hippocampal formation (Hipp, Sb, pSb).

Some of the rostral intralaminar nuclei also appear to have a

close association with components of the fronto-limbic complex,

in contrast to the caudal intralaminar nuclei (see below).

Midline and Intralaminar Nuclei

Figures 1 and 2 suggest that extrinsic connectivity defines two

broad categories of midline and intralaminar nuclei. In black

(Figs 1 and 2) are CLN, CMN and PAC (of the rostral intralaminar

group, see Appendix) and PF and CM (of the caudal intralaminar

group, Appendix 1). These nuclei connect very widely with most

areas of the visual, auditory, somato-motor and the fronto-limbic

systems. In brown, the same colour as the fronto-limbic

complex, are MV-RE, PAT, PARA and PARP (of the caudal intra-

laminar group, Appendix 1). These nuclei make few connections

outside the fronto-limbic complex. This connectional difference

suggests a distinction between CLN, CMN, PAC, CM and PF on

the one hand, and RH, MV-RE, PAT, PARA and PARP on the other.

CLN, CMN, PAC, CM and PF correspond to the ‘intralaminar

complex’ proposed by Berman and Jones (1982). These analyses

agree with Berman and Jones’ suggestion (1982) that RH, MV-RE,

PAT, PARA and PARP do not necessarily constitute a functional

group.

Optimal Set Analysis

We can be confident that optimal set analysis found genuine sets

within the connection data because the low costs of the solu-

tions when compared with solutions computed from randomly

shuff led connection matrices. For example, at balanced

attraction and repulsion (Figs 8 and 9) with real connection data,

the mean (± SEM) cost of the first 20 generations of the first

epoch was 1861 15. The mean of the equivalent cost calculated

from five randomly shuff led connection matrices was 2729, with

a SD of 42. Therefore, in this case, the cost of sets calculated

from connection data was ∼20 z-scores less than the mean cost of

sets calculated with random data (P < 10–12). In general, the costs

of solutions computed from connection data were broadly

comparable with the costs of solutions computed from four- to

286 Cat Cortico-thalamic Network • Scannell et al.

Figure 3. Optimal set analysis at balanced attraction and repulsion. The figure summarizes the 31 solutions whose costs lie within 1% of the single lowest cost solution. Attraction(the factor multiplied by connection weight to compute the cost of connections between sets) was 1 and repulsion (the cost of non-connections within sets) was 1. Cortical areasand thalamic nuclei are listed along the top and left of the table. The frequency with which areas or nuclei co-localized within the same set is given by the colour (scale at top rightof table). Darker colours show frequent co-localization, while light colours show rare co-localization. The analysis consistently finds a number of sets. From top to bottom of Figure 3,these are: a set of frontolimbic, multimodal and higher-order somato-motor structures (LM-Sg to PFCMil); a set of auditory structures (MGM to MGDd); a set of lower-ordersomato-motor structures (4g to 3b); two small fronto-limbic sets (Sb to Hipp, and pSb to RS); and a large visual set (LPl to PMLS). Some structures do not have a strong affiliationwith any of the large sets, either because of few known connections (e.g. LGLGW) or because of strong connections with more than one set (e.g. 7 and ALLS). The number of setschanges substantially when the attraction and repulsion parameters are at different values.

Cerebral Cortex Apr/May 1999, V 9 N 3 287

six-dimensional test data (Scannell et al., 1997; J.W. Scannell,

unpublished observations).

We performed optimal set analysis with attraction weights

ranging from 1 to 10, with repulsion at 1, and with repulsion

weights ranging from 2 to 30, with attraction at 1. For each of

the levels of attraction and repulsion, we ran CANTOR for 50

epochs from 50 different random starting configurations. At any

given level of attraction and repulsion, optimal set analysis tends

to produce a lot of solutions with similar low costs and there is

no principled way of choosing which is ‘best’.

Figure 3 shows a summary of the 31 solutions we found with

costs within 1% of the single lowest-cost solution at balanced

attraction and repulsion. The balanced attraction and repulsion

case is the most intuitively obvious. In this condition, the cost of

a connection between sets is equal to the weight of the

connection (1, 2 or 3) and the cost of a non-connection within

sets is equal to 1. The difference between the 31 solutions is

small and all agree on several clear and consistent sets (solid dark

blocks in Fig. 3). From top to bottom of Figure 3, these

correspond to a set containing higher somato-motor and some

multimodal and fronto-limbic structures (LM-Sg to PFCMil); an

auditory set (MGM to MGDd); a set of lower somato-motor

structures (4g to 3b); two small sets of frontolimbic structures

(one consisting of Hipp, Sb and MV-Re, and the other containing

pSb, LD, AD, AV and RH); and a large set of visual structures (LPl

to PS). Around 17 areas or nuclei have variable positions and do

not consistently lie in any of these sets. Some of these structures

have few connections, so are ‘pushed’ out of sets by the repulsive

effect of many non-connections (e.g. LGLA, LGLC). Other areas

and nuclei may have substantial connections with structures in

more than one of the major sets (e.g. 7, CMN, 36, Enr, POm, SG),

so fit into different sets with very little effect on the overall cost

of the configuration.

Optimal set analysis is in excellent agreement with NMDS

over the main features of the connectional architecture of the

cat thalamo-cortico-cortical network. Figure 4 plots the major

Figure 4. Optimal set analysis and NMDS are in excellent agreement. The figure is a stereogram that shows the major sets identified by optimal set analysis (Fig. 3) colour-codedand plotted on three-dimensional MDS configurations (Fig. 1). The set of fronto-limbic, multimodal, and higher-order somato-motor structures (LM-Sg to PFCMil in Fig. 3) is colouredbrown; the set of auditory structures (MGM to MGDd in Fig. 3) is coloured green; the set of lower order somato-motor structures (4g to 3b in Fig. 3) is coloured red; The ‘hippocampal’set (Sb to Hipp in Fig. 3) is coloured purple; the ‘anterior nuclei’ set (LD to RS in Fig. 3) is coloured blue/green; and the visual set (LPl to PMLS in Fig. 3) is coloured blue. The structuresoutside these sets are black. The figure was computed using all the connection data but, for clarity, only strong reciprocal connections, with a sum of weights in both directions >4,are drawn. The spatial configurations of points are exactly as in Figure 1. The sets identified by optimal set analysis map neatly onto compact regions of space identified by NMDS.Furthermore, the unclassified (black) structures tend to occupy peripheral regions, or are near the borders between the sets. This shows that NMDS and optimal set analysis agreeover the major features of the connectional organization of the thalamo-cortical network.

288 Cat Cortico-thalamic Network • Scannell et al.

Figure 5. Summary of optimal set analysis with the attraction parameter ranging from 1 to 10 (repulsion set to 1), and repulsion parameter ranging from 2 to 30 (attraction set to1). Cortical areas and thalamic nuclei are listed along the top and left of the table. The frequency with which areas or nuclei co-localized within the same set is given by the colour(scale at top right of table). Darker colours show frequent co-localization within the same set, while light colours show rare co-localization.

Cerebral Cortex Apr/May 1999, V 9 N 3 289

Figure 6. Cluster structure of the cat thalamo-cortical network. The figure shows the relative frequency with which cortical areas occupied the same cluster in 10 clusteringparadigms applied to each of six alternative five-dimensional NMDS configurations (see Materials and Methods). Pooling results from a wide range of clustering and NMDS methodsaimed to average out the peculiarities of individual analysis methods, revealing robust connectional features. Dark squares represent areas that frequently occupied the same cluster,while light squares represent areas that are rarely occupied the same cluster (key on right of figure). The names of anatomical structures are listed down the left margin and acrossthe top of the matrix. To ease comparison with Figure 3, the gross order of areas is comparable to that in Figure 3, although the order of areas has been changed locally (within thebasic ‘sets’ identified in Fig. 3) to illustrate additional features.

290 Cat Cortico-thalamic Network • Scannell et al.

sets illustrated in Figure 3 on a stereogram showing three-

dimensional configurations yielded by NMDS. Figure 4 shows

that the sets (Fig. 3) map neatly and compactly onto the

three-dimensional NMDS configurations with little stretching,

fracturing or overlap between the sets. It is clear that most of the

unclassified structures lie either at the edge of the NMDS

configurations (those with few connections), or else lie near the

borders between the sets (those with substantial connections

with more than one set). Thus there is excellent agreement

between two entirely independent data analysis methods.

Over the full range of attraction and repulsion parameters,

high attraction gave fewer, larger, clusters, identifying higher-

order associations between the major thalamo-cortical systems.

High repulsion gave more, smaller, clusters, identifying the most

closely associated areas or nuclei. Figure 5 shows the pooled

results from the battery of attraction and repulsion levels.

Figures 5 and 3 are in good agreement, but the graded nature of

Figure 5 reveals both the local and the large-scale structure of the

thalamo-cortical network and allows one to look across levels

of organization. For example, the darkest squares show areas or

nuclei that very frequently co-localize across all levels of

repulsion so are therefore very strongly associated. The lightest

squares show areas or nuclei that very rarely co-localize, even

at the highest levels of attraction, so are therefore very con-

nectionally distant.

In line with the balanced attraction and repulsion case (Figs 3

and 4), the somato-motor system is closely associated with

fronto-limbic multimodal cortical areas. The first large block of

structures in Figure 5 (6l to MV-Re) contains fronto-limbic,

somato-motor and multimodal structures. Figure 5 confirms that

the auditory system (MGV to AII) is the most topologically

isolated of the thalamo-cortical systems, although MGV and Tem

occasionally co-localize with fronto-limbic and visual structures.

Within the visual system (LGLA to LGLGW), ALLS and PLLS

sometimes co-localize with AES and IG, and LM-Sg. 5Bl is closely

associated with many visual structures.

A feature of optimal set analysis applied to this data set is that

areas or nuclei with few reported connections often end up in

sets on their own (e.g. LIc, Hipp). Some areas or nuclei have few

reported connections because they genuinely make few, possibly

very strong, connections (e.g. LGLA, LGLC). Others have few

reported connections because of lack of study or lack of

consensus over their existence or classification (possibly LIc, and

PF). These aspects of the analyses would benefit from more

experiments on the less studied areas and from quantitative

connection data for all areas. Quantitative data, which would

provide a more accurate cost function, is provided by some

researchers (e.g. Sherk, 1986; Musil and Olson, 1991; Peters et

al., 1993, 1994; MacNeil et al., 1997; Van Duffel et al., 1997), but

is unavailable for the vast majority of connections reported in the

neuroanatomical literature.

Cluster Analysis

Figure 6 shows the pooled results from the battery of non-

parametric cluster analysis methods. The areas in Figure 6 have

been ordered so that the sets are comparable to Figure 3. It is

clear that the cluster analysis methods are in good agreement

with optimal set analysis (Figs 3 and 5). As Figure 6 shows pooled

results from a range of NMDS methods, and a range of clustering

paradigms with a widely differing kernel radii, it has a graded

nature that reveals considerable detail about both the local and

the large-scale structure of the thalamo-cortical network.

Large-scale structure has been identified by analyses using large

radii kernels and the local structure has been revealed by the

analyses with small radii kernels.

The ability of the analysis to look across different scales means

that Figure 6 contains a wealth of information that cannot all be

explored in detail here. However, we outline several example

features to help the reader to explore the Figure. First, the region

corresponding to the first set from Figure 3 (containing higher

somato-motor and some multimodal and fronto-limbic struc-

tures) is clearly not homogenous. The areas of prefrontal cortex

are much more closely associated with each other than, for

example, with VA-VL, 6m, 6l and 7. Second, VA-VL, 6m, 6l, 7, 5Al

and 5Bl have associations with structures that lie within the ‘low

somatomotor’ set of Figure 3, namely 4g, 4, 5Am, SSAi, SSAo and

5Bm. Third, the region of Figure 6 corresponding to the visual

set of Figure 3 also shows considerable substructure. For

example, there are particularly close connectional associations

between 17 and 18, and between DLS, LPm and PLLS. These

three structures also have associations outside the visual set,

with AES and ALLS.

Figure 6 shows several distinct and robust small connectional

groups that are relatively invariant with NMDS method and

clustering paradigm (darkest regions in Fig. 6). These groups

must contain structures that are in very close connectional

association since they are consistently identified in different

NMDS configurations even by clustering methods with small

kernels. From top to bottom of Figure 6, particularly tight

subgroups include PFCMil, PFCMd and PFCL; CGa and CGp; 6l

and 6m; 5Al, 5Bl and 7; AES, SIV and ALLS; 35 and 36; AAF and

AI; Tem and MGvl; a low somato-motor group including 1, 2, 3a,

3b, SII, and VPL and VPM; Sb and MV-Re; areas 17 and 18; and

LPm and PLLS.

An alternative and informative reading of Figure 6 is to

concentrate on the very light regions which show those groups

of areas that are least likely to occur within the same clusters.

The unimodal visual structures (LGLGW to PLLS) are very distant

from most auditory structures (MGV to MGDd), and from the

early somato-motor structures (4g to SII). The early somato-

motor structures (4g to SII) are very distant from the early

auditory system (MGV to MGDd). All the early sensory and

sensory-motor groups are more distant to each other than they

are to the fronto-limbic and higher order sensory-motor regions

(e.g. PFCL to 36, and PAT to RS). This is consistent with the

notion that, at a gross level of description, the thalamo-cortical

network is organized into three distinct sensory or sensory

motor systems and a topologically central group of structures. In

this analysis, the topologically central group of structures

includes fronto-limbic structures, some higher somato-motor

(e.g. 6l, 5m) and possibly multimodal sensory structures (e.g.

AES, 7).

Discussion

A New Approach

This paper presents an approach to the analysis of anatomical

connection data that is still relatively new and that frequently

invites questions. First, can one collate brain connectivity data

from many different studies on many different individuals?

Second, is it practical to produce simple representations of such

complicated data? Third, what are the advantages and

disadvantages of the different analysis methods? Fourth, what are

the limitations of the ‘grey box’ level of explanation (Douglas

and Martin, 1991) that considers cortical areas, thalamic nuclei

and extrinsic connections as the basic units of organization?

Cerebral Cortex Apr/May 1999, V 9 N 3 291

Collation and Variability

There are several reasons to collate anatomical connection data.

First, collation may highlight gaps in current knowledge (e.g.

cortical connections to the midline and intralaminar nuclei, see

Table 2 and Results section; see also Hilgetag et al., 1996).

Second, it may direct further empirical study (e.g. Scannell et al.,

1996). Third, collation provides the opportunity to test theories

of neuroanatomical organization with a degree of quantitative

rigor (e.g. Cherniak, 1990; Nicolelis et al., 1990; Musil and

Olson, 1991; Young, 1992; Hilgetag et al., 1996; Scannell, 1997).

Fourth, primary connection data are distributed across a large

and, to the non-specialist, difficult literature. Collation can allow

a wide range of researchers to benefit from advances in

experimental neuroanatomy.

Despite these advantages, collation is complicated by varia-

bility in anatomists’ parcellation schemes, connection-tracing

methods, and by variability in the results of equivalent tracer

deposits in different animals. Variability has been highlighted in

recent quantitative studies and is substantial both in anatomical

structures (e.g. Rajkowska and Goldman-Rakic, 1995; Andrews

et al., 1997), and in the density of connections between them

(MacNeil et al., 1997). MacNeil et al. (1997) have counted the

number of labelled thalamic and cortical neurons following

injections in the medial suprasylvian visual area. They found that

visual cortico-cortical connection densities vary widely between

tracer deposits. At first sight, high variability (whether within or

between individuals) seems problematic for much of connec-

tional neuroanatomy, and particularly so for studies such as this

one.

MacNeil et al.’s (1997) results have prompted us to examine

quantitative data from Musil and Olson (1988), Bowman and

Olson (1988), Olson and Jeffers (1987), Olson and Lawler

(1987), Olson and Musil (1992), and Musil and Olson (1991),

and unpublished quantitative connection data on the lateral

suprasylvian areas (S. Grant, personal communication). The data

suggest that for a wide range of cortico-cortical and thalamo-

cortical connections, the distribution of connection densities

across equivalent experiments in different individuals is highly

non-normal. In fact, the distribution is closer to an exponential

(R.J. Baddeley and J.W. Scannell, unpublished observations).

This has important implications. First, statistical methods based

on the assumption of a normal distribution are not appropriate

for the analysis of such data. Second, the exponential distribu-

tion is extremely variable, so large amounts of data are required

for confident estimates of sample statistics. Third, the standard

measure of variability (sample SD) can show a relatively strong

bias when applied to exponentially distributed data. For small

sample sizes, the variability will be systematically under-

estimated. Table 4 (R.J. Baddeley and J.W. Scannell, unpublished

observations) provides a guide to confidence intervals for

estimates of mean and SD connection density. It also provides

a guide to correcting the bias in sample SD. By providing

confidence intervals, a simple indicator of statistical power,

Table 4 should help researchers to choose appropriate sample

sizes in connection tracing experiments.

Table 4 shows that small samples (five or fewer individuals)

have very wide confidence intervals, so are likely to suffer from

serious random sampling error. The benefits of increasing

sample size begin to tail off with sample sizes of 10 or more

individuals (Table 4). This suggests that studies of the con-

nections to a particular cortical area should aim to inject the area

in ∼10 individuals if reasonable quantitative (or qualitative)

estimates of connection densities are required.

Many primary studies of cortico-cortical and cortico-thalamic

connections are based on substantially fewer than 10 cases, so

risk serious random sampling error. However, where several

primary studies have examined the same connection, collation is

likely to reduce the problem of random sampling error in the

estimation of connection density. This gives some justification to

those who have made serious attempts to distil data from many

neuroanatomy papers into a single coherent and analysable form

(e.g. Pandya and Yeterian, 1985; Felleman and Van Essen, 1991).

However, with high variability, presenting any single con-

nectional scheme as representative becomes difficult. This is

because most individuals depart considerably from the ‘average’

pattern of connection densities. Indeed, no single individual can

have a pattern of connection densities that is very representative

of the others. To ref lect this genuine aspect of connection data,

the nature of the distribution of connection densities should be

quantified (Cherniak, 1990), reported and represented in future

attempts at collation and modelling (MacNeil, 1997).

Complexity

The second question, that connection data are too complicated

to summarize, is serious and plausible, but, our data suggest,

untrue. One can imagine a brain in which the pattern of

area-to-area connections was so complex that it could be

described by nothing simpler than the connection matrix itself.

Putting random numbers in Table 2 would simulate the

complexity of connectivity of such a brain. Fortunately, the

goodness of fit of the NMDS configurations and the low cost of

the CANTOR solutions, when compared with analyses of test

data, show that the connectional architecture of the real

thalamo-cortical network can be well represented relatively

simply. This finding is important. It suggests that, in principle,

relatively few sources of variability can account for much of the

Table 4Guide to sampling in quantitative connection tracing experiments

N (A) Mean (B) Standard deviation

Upper Lower Upper Bias Lower

95% 70% 70% 95% 95% 70% 70% 95%

2 2.83 1.70 0.54 0.12 2.59 1.36 0.71 0.25 0.024 2.20 1.50 0.69 0.28 2.24 1.34 0.84 0.50 0.156 1.93 1.43 0.75 0.37 2.04 1.33 0.88 0.59 0.268 1.80 1.37 0.79 0.44 1.89 1.32 0.90 0.65 0.31

10 1.69 1.31 0.81 0.49 1.82 1.30 0.92 0.69 0.3712 1.63 1.30 0.83 0.52 1.79 1.29 0.93 0.72 0.4014 1.58 1.27 0.85 0.55 1.74 1.27 0.94 0.75 0.4316 1.54 1.26 0.86 0.57 1.68 1.25 0.95 0.76 0.4618 1.50 1.25 0.87 0.59 1.66 1.24 0.95 0.77 0.4920 1.48 1.24 0.87 0.61 1.62 1.24 0.95 0.79 0.51

The table provides a guide to sampling from the exponential distribution, which bears a closerresemblance to the quantitative distribution of connection densities than does the normaldistribution (R.J. Baddeley and J.W. Scannell, unpublished observations). The left column, N,shows sample size (the number of individuals with tracer deposits). 70% confidence limits areanalogous to standard error. (A) Mean. Columns show the ratio of the upper and lower 95% and70% confidence limits of the population mean to the sample mean. These are proportional tosample mean for data from an exponential distribution. Therefore, the intervals for any givensample can be calculated by multiplying sample mean by the appropriate confidence limits. Forexample, with a sample size of six and a sample mean connection density of 1000 labelledneurons, the 95% confidence interval for the population mean ranges from 370 to 1930 labelledneurons. The 70% confidence interval ranges from 750 to 1430 neurons. (B) Standard deviation.Columns show the ratio of the upper and lower 95% and 70% confidence limits of population SDto uncorrected sample SD. The ‘bias’ column shows the expected ratio of sample SD to populationSD. Unbiased estimates of population standard deviation may be calculated by dividing samplestandard deviation by the appropriate bias. For example, with a sample size of four and a sampleSD of 1000 neurons, the unbiased estimate of population SD is 1192 (which equals 1000/0.84),and the 95% confidence interval ranges from 150 to 2240 neurons.

292 Cat Cortico-thalamic Network • Scannell et al.

variability in the gross connection pattern of the thalamo-cortical

network. We think this apparent simplicity shows that a small

number of simple rules, ref lecting genetic and/or developmental

economy, have been favoured by evolution in the design of the

thalamo-cortical network.

Analysis Methods

Many anatomists have sought to present complex connection

data in a compact and comprehensible way. McCulloch (1944)

provides an excellent example. However, the complexity of

modern data make the problem considerable (Nicolelis et al.,

1990; Felleman and Van Essen, 1991). While our methods make

a contribution, none are an ideal solution.

NMDS works with ordinal data to give graphical repres-

entations of connectivity. However, there are many possible

NMDS algorithms and cost functions and it is not always clear

which, if any, is best. We are also limited to presenting

configurations in no more than three dimensions. We think this

acceptable for the thalamo-cortical data in this paper, but it may

not be appropriate for other anatomical data. Data that cannot be

represented by a low-dimensional spatial arrangement of points

(e.g. data with many one way connections, where the proximity

of A to B does not equal the proximity of B to A) are unsuitable

for graphical presentation with NMDS.

Optimal set analysis avoids dimensional reduction and allows

us to specify particular cost functions; in this case, functions

that embody an attractive definition of a neural system. Varying

the cost function allows us to look for both local and global

features of neural organization. However, the method tends to

produce a very large number of optimal solutions that then have

to be summarized (e.g. Hilgetag et al., 1996). Sometimes the

summaries are simple (e.g. Fig. 3), but they may be nearly as

complex as the connection matrix itself (e.g. Fig. 5). In addition,

our cost function assumes a metric relationship between

connection densities, which is a disadvantage when compared

with NMDS.

Non-parametric cluster analysis compromises between NMDS

and optimal set analysis. It balances the need to avoid dimen-

sional reduction with the need to avoid the problems of density

estimation in high-dimensional spaces (Silvermann, 1986; Burns,

1997) by using five-dimensional NMDS configurations as a

starting point. The method can look for both local and global

organization. However, the method is less transparent than

either NMDS or optimal set analysis and we have to perform

many different cluster analyses and pool the pool the results, so

that spurious clusters (Gordon, 1996) are averaged out. This

makes the summary data complicated.

Despite the limitations of the different methods, we are

encouraged by the fact that they yield largely consistent results.

We hope they provide relatively comprehensible working

models which can, at best, serve as rough ‘route planners’ to the

large-scale organization of the thalamo-cortical network.

Levels of Analysis

The final question relates to the level of analysis presented in this

paper. Is it useful to pursue the somewhat abstract level of

analysis in which the cortical area, thalamic nucleus and

extrinsic connection are the basic units?

Our choice of level of analysis has been pragmatic and

utilitarian, seeking the greatest number of connections for

the greatest number of structures. We have tried to ref lect the

cortical areas and thalamic nuclei of the cat that have been

established by other authors and for which there are connection

data. However, the distinctions we have drawn are not neces-

sarily the most principled. For example, we have subdivided the

LGN into several subnuclei on the basis of morphological,

connectional and physiological differences between neuronal

populations, but have treated area 17 as a single unit. However,

each layer of area 17 has different extrinsic connections,

physiological properties and morphology. It may also be the case

that cortical areas and thalamic nuclei are not an appropriate

level of organization for our kind of analysis of connectivity. For

example, in the early visual system, parts of different areas

representing similar regions of retinotopic space may be much

more functionally and connectionally associated than parts of

the same area representing different regions of space.

We see no easy answer to these questions. However, in the

long term, quantitative data on connectivity, morphology and

physiology are necessary to determine the most natural

parcellation schemes and levels of analysis.

Indeterminate Hierarchies and Directed Loops

Crick and Koch (1998) suggest that directed feed-forward loops

should not exist within the extrinsic thalamo-cortico-cortical

network, as they would cause runaway excitation (Crick and

Koch, 1998). They also suggest that by incorporating thalamo-

cortical connections, ‘indeterminate’ hierarchies (Hilgetag et al.,

1996) might be avoided.

If the direction of connections is only known at the

categorical level of measurement (‘ascending’, descending’ or

‘lateral’), thalamo-cortical connections cannot remove indeter-

minacy from hierarchical analysis. Cortical hierarchies (e.g.

Felleman and Van Essen, 1991) are indeterminate (Hilgetag et al.,

1996) because connectivity is too sparse to constrain a unique

hierarchy, even if the direction of every cortical connection

were known. Indeterminacy will not go away if thalamo-cortical

connections are included. This is because the cortico-thalamo-

cortical loops that could, in principle, provide additional

hierarchical information do, in fact, only link cortical areas that

already have direct cortical projections (Scannell et al., 1997).

The exceptions to this rule are connections to and from the

midline or intralaminar nuclei (Scannell et al., 1997). The lack of

directional thalamic ‘shortcuts’ between areas without direct

cortical connections means that thalamic connections cannot

provide the additional information necessary remove indeter-

minacy from sensory-motor hierarchies (cf. Crick and Koch,

1998). Indeterminacy may disappear if it proves possible to

obtain metric measures of the direction of cortical connections,

as suggested by recent work of Barbas and Rempel-Clower

(1997).

Symmonds and Rosenquist (1984) and Scannell et al. (1995)

noted that the origin and termination patterns of connections

between cat cortical areas frequently do not match the ‘feed-

forward’, ‘feed-back’ and ‘lateral’ templates that have been

applied in the primate (e.g. Felleman and Van Essen, 1991; Crick

and Koch, 1998). For example, cat area 17 sends projections that

originate predominantly in the layers II and III (Rosenquist,

1985) and terminate predominantly in layers II and III of areas 18

and 19 (Price and Zumbroich, 1989). The projection from area

18 to 17 originates in layers II, III and V and terminates in layers

I, II, III, IVA, V and superficial layer IV of area 17, while tending

to avoiding deep layer VI and layer IVB (Henry et al., 1991). At

present, the functional implications of the differences in laminar

origin and termination patterns in the cat and macaque cortices

are unclear. However, the fact that cats are frequently used for

studies of cortical microcircuitry, while macaques are favoured

Cerebral Cortex Apr/May 1999, V 9 N 3 293

for systems neuroscience, means that these differences require

further critical evaluation.

Structure–Function Relationships

Connectional neuroanatomy is driven by the idea that there is a

close link between brain connectivity and brain function (e.g.

Meynert, 1890). Van Duffel et al. (1997) have recently explored

this relationship in an ingenious experiment which correlated

quantitative measures of anatomic and metabolic connectivity.

However, we believe that the efforts of anatomists in finding and

describing connections have not been efficiently exploited to

explore the structure–function relationship. This is because

functional arguments based on extrinsic connectivity have

tended to be informal, non-quantitative and/or highly simplistic.

Recently, Hilgetag et al. (1997a) have developed more formal

network models, whose connectivity is based on the pattern of

extrinsic cortical connections collated by Scannell et al. (1995),

Young (1993) and others. These models provide excellent

quantitative accounts of the spread of electrical activity found in

old strychnine neuronographic experiments. It may be possible

to extended such methods to explore the functional con-

sequences of cortical lesions. Since the properties of networks

may not be intuitively obvious, and since neuronal activity can

spread via direct and indirect connections, simple models

relying only on the strength of direct connections (e.g. Van

Duffel et al., 1997) may not provide a realistic account of

functional connectivity within the thalamo-cortical network.

Efficient Wiring

It has been suggested (Cherniak, 1992, 1994, 1995, 1996) that

the organization of the central nervous system can be explained

if ‘saving wire’ is the primary and overriding design constraint.

While efficient wiring has long been thought to contribute

to some aspects of brain organization (e.g. Cowey, 1984;

Mitchison, 1991; Young, 1992; Cherniak, 1994; Scannell et al.,

1995; Scannell, 1997; Van Essen, 1997), Young and Scannell

(1996) have argued that many other constraints are also

important. If saving wire were the overriding design constraint,

connectional proximity should map very neatly onto physical

proximity. Component placement optimization, therefore,

predicts a physical brain organization that resembles Figure 1. To

‘save wire’, thalamic nuclei should dissociate from their

neighbours, with which they do not connect, and migrate to be

close to the cortical areas with which they do connect. However,

the thalamic nuclei remain firmly clustered together near the

centre of the cerebral hemisphere. This may ref lect develop-

mental and evolutionary constraints. The thalamus was well

developed in reptiles, long before the origin of neocortex

(Masterton, 1976). Furthermore, the development of the

thalamus and thalamo-cortical projections precedes that of the

cortical plate and cortico-cortical projections (O’Leary et al.,

1994; Payne et al., 1988). Therefore, constraints other than the

need to save wire may strongly constrain the organization of the

central nervous system (Young and Scannell, 1996; Scannell,

1997).

Global Architecture

Young (1993), Scannell and Young (1993) and Young et al.

(1995a,b) found that the extrinsic connectional architecture of

the cortico-cortical network of macaques and cats defined four

major systems. There were the visual, auditory and somato-

motor systems and a fourth system, termed the ‘fronto-limbic’

system, composed of prefrontal cortex and structures in the

medial part of the hemisphere (sometimes termed ‘limbic’

cortex). The sensory-motor systems all showed a degree of

hierarchical organization and their highest stations were in close

connectional association with the fronto-limbic complex that

therefore has access to the most highly processed sensory and

motor information. There was more limited cross-talk between

the systems outside the fronto-limbic complex, but this differed

between the cat and macaque.

Our results show that a similar plan holds for the thalamo-

cortico-cortical network. Different analysis methods consistently

detect three groups of areas corresponding to the unimodal

components of the visual, auditory and somato-motor systems.

The higher stations of these systems are associated with a

topologically central group of structures (insular, cingulate and

prefrontal cortex, associated nuclei and, depending on analysis

method, the higher stations of the somato-motor system and

multimodal cortical areas). The analyses also identify a final

group of structures (e.g. Hipp, MV-Re, Sb, pSb, AM, AD, Enr),

that are at the furthest possible remove from the sensory-motor

periphery. This basic plan is a robust and reproducible feature of

three independent analyses of a comprehensive body of con-

nection data. Therefore, it should represent a strong constraint

on large-scale theories of the organization of the brain.

NotesThis work was supported by the Wellcome Trust.

Address correspondence to J.W. Scannell, Neural Systems Group,

Psychology Department, Ridley Building, University of Newcastle,

Newcastle upon Tyne NE1 7RU, UK. Email: J.W.Scannell@ncl.ac.uk.

ReferencesAndersen RA, Knight PL, Merzenich MM (1980) The thalamocortical and

corticothalamic connections of AI, AII, and the anterior auditory field

(A AF) in the cat: evidence for two largely segregated systems of

connections. J Comp Neurol 194:663–701.

Andrews TJ, Halpern SD, Purves D (1997) Correlated size variations in

human visual cortex, lateral geniculate nucleus, and optic tract.

J Neurosci 17:2859–2868.

Asuncion-Moran M, Reinoso-Suarez F (1988) Topographical organization

of the thalamic afferent connection to the motor cortex in the cat.

J Comp Neurol 270:64–85.

Avendano C, Rausell E, Reinoso-Suarez F (1985) Thalamic projections to

areas 5a and 5b of the parietal cortex in the cat. A retrograde

horseradish peroxidase study. J Neurophysiol 5:1446–1470.

Avandano C, Stepniewska I, Rausell E, Reinoso-Suarez F (1990)

Segregation and heterogeneity of thalamic cell populations projecting

to superficial layers of posterior parietal cortex: a retrograde trace

study in cat and monkey. Neuroscience 39:547–559.

Barbaresi P, Minelli, A, Manzoni T (1991) Thalamic connections of the

second somatic sensory area in cats studied with anterograde and

retrograde tract-tracing techniques. Neuroscience 46:159–163.

Barbas, H, Rempel-Clower N (1997) Cortical structure predicts the

pattern of corticocortical connections. Cereb Cortex 7:635–646.

Bentivoglio M, Albanese A, Macchi G (1981) The cortical projections of

the thalamic intralaminar nuclei, as studied in cat and rat with the

multiple f luorescent retrograde tracing technique. Neurosci Lett

26:5–10.

Berman AL, Jones EG (1982) The thalamus and basal telencephalon of the

cat. Madison: University of Wisconsin Press.

Berson DM, Graybiel AM (1991) Tectorecipient zone of the cat lateral

posterior nucleus: evidence that collicular afferents contain acety-

cholinesterase. Exp Brain Res 84:478–486.

Birnbacher D, Albus K (1987) Divergence of single axons in afferent

projections to the cat’s visual cortical areas 17, 18 and 19: a

parametric study. J Comp Neurol 261:543–561.

Bishop CM (1995) Neural networks for pattern recognition. Oxford:

Clarendon Press.

Bowman EM, Olson CR (1988) Visual and auditory association areas of the

cats’ posterior ectosylvian gyrus — thalamic afferents. J Comp Neurol

272:15–29.

294 Cat Cortico-thalamic Network • Scannell et al.

Brandner S, Redies H (1990) The projection from the medial geniculate to

field AI in the cat: organization in the isofrequency dimension.

J Neurosci 10:50–61.

Brugge JF, Reale RA (1985) Auditory cortex. In: Cerebral cortex, vol. 4

(Peters AJ, Jones EG, eds). New York: Plenum Press.

Bullier J, Kennedy H, Salinger W (1984) Branching and laminar origin of

projections between visual cortical areas in the cat. J Comp Neurol

228:329–341.

Burns GAPC (1997) Neural connectivity of the rat: theory, methods and

applications. D.Phil. thesis, Oxford University.

Burton H, Kopf EM (1984) Connections between the thalamus and the

somatosensory areas of the anterior ectosylvian gyrus in the cat. J

Comp Neurol 224:173–205.

Calford MB (1983) The parcellation of the medial geniculate body of the

cat defined by the auditory response properties of single units. J

Neurosci 3:2350–2364.

Calford MB, Aitkin LA (1983) Ascending projections to the medial

geniculate body of the cat: evidence for multiple parallel pathways

through the thalamus. J Neurosci 3:2365–2380.

Cavada CR-SF (1985) Topographical organization of the cortical afferent

connections of the prefrontal cortex. J Comp Neurol 242:293–324.

Cherniak CC (1990) The bounded brain: towards quantitative neuro-

anatomy. J Cogn Neurosci 2:58–68.

Cherniak CC (1992) Local optimization of neuronal arbors. Biol Cybernet

66:503–510.

Cherniak CC (1994) Component placement optimization and the brain.

J Neurosci 14:2418–2427.

Cherniak CC (1995) Neural component placement. Trends Neurosci

18:522–527.

Cherniak CC (1996) Component placement optimization in the brain — a

reply. Trends Neurosci 19:414–415.

Clarey JC, Irvine D (1990) The anterior ectosylvian sulcal auditory field in

the cat. 2.A horseradish-peroxidase study of its thalamic and cortical

connections. J Comp Neurol 301:304–324.

Clasca F, Llamas A, Reinoso-Suarez F (1997) Insular cortex and

neighbouring fields in the cat: a redefinition based on cortical

microarchitecture and connections with the thalamus. J Comp Neurol

384:456–482.

Colby CL, Duhamel J (1991). Heterogeneity of extrastriate visual areas

and multiple parietal areas in the macaque monkey. Neuropsychologia

29:417–537.

Contreras D, Destexhe A, Sejnowski TJ, Steriade M (1996) Control of

spatiotemporal coherence of a thalamic oscillation by cortico-

thalamic feedback. Science 274:771–774.

Cowey A (1979) Cortical maps and visual perception. Quart J Exp Psychol

31:1–17.

Craig AD, Wiegand SJ, Price JL (1982) The thalamo-cortical projection of

the nucleus submedius in the cat. J Comp Neurol 206:28–48.

Crick F, Koch C (1998) Constraints on cortical and thalamic projections:

the no-strong-loops hypothesis. Nature 391:245–250.

Cunningham ET, Levay S (1986) Laminar and synaptic organization of the

projection from the thalamic nucleus centralis to the primary visual

cortex in the cat. J Comp Neurol 254:65–77.

Douglas RJ, Martin KAC (1991) Opening the gray box. Trends Neurosci

14:286–293.

Dreher B (1986) Thalamocortical and corticocortical interconnections in

the cat visual system: relation to the mechanisms of information

processing. In: Visual neuroscience (Pettigrew JD, Sanderson KJ,

Levick WR, eds), pp. 290–314. Cambridge: Cambridge University

Press.

Dykes RW, Herron P, Lin CH (1986) Ventroposterior thalamic regions

projecting to cytoarchitectonic areas 3a and 3b in the cat. J

Neurophysiol 56:1521–1541.

Edgington ES (1980) Randomization tests. New York: Marcel Dekker.

Epanechnikov VA (1969) Nonparametric estimation of a multi-

dimensional probability density. Theory Prob Appl 14:153–158.

Feig S, Harting JK (1998) Corticocortical communication via the

thalamus: ultrastructural studies of corticothalamic projections from

area 17 to the lateral posterior nucleus of the cat and inferior pulvinar

nucleus of the owl monkey. J Comp Neurol 395:281–195.

Felleman DJ, Van Essen DC (1991) Distributed hierarchical processing in

the primate cerebral cortex. Cereb Cortex 1:1–47.

Felleman DJ, Wall JT, Cusick CG, Kaas JH (1983) The representation of the

body surface in S-I of cats. J Neurosci 3:1648–1669.

Fujii M (1983) Fiber connections between the thalamic posterior lateral

nucleus and the cingulate gyrus in the cat. Neurosci Lett 39:137–142.

Geisert EE (1980) Cortical projections of the lateral geniculate nucleus in

the cat. J Comp Neurol 190:793–812.

Glenn LL, Hada J, Roy JP, Deschenes M, Steriade M (1982) Anterograde

tracer and field potential analysis of the neocortical layer 1 projections

from nucleus ventralis medialis of the thalamus of the cat.

Neuroscience 7:1861–1877.

Gomez-Pinilla F, Villablanca JR (1989) Motor cortex projections to the

thalamus. An autoradiographic study in the cat. Arch Ital Biol

127:109–144.

Goodhill GJ, Simmen MW, Willshaw DJ (1995) An evaluation of the use of

multidimensional scaling for understanding brain connectivity. Phil

Trans R Soc. B 348:265–280.

Gordon AD (1996) Cluster validation. In: Fifth Conference of the

International Federation of Classification Societies. Kobe, Japan.

Gower JC (1971) Statistical methods of comparing different multivariate

analyses of the same data. In: Mathematics in the archeological and

historical sciences, pp. 138–149. Edinburgh: Edinburgh University

Press.

Grant S, Shipp S (1991) Visuotopic organization of the lateral suprasylvian

area and of an adjacent area of the ectosylvian gyrus of cat cortex — a

physiological and connectional study. Vis Neurosci 6:315–338.

Graybiel AM, Berson DM (1980) Histochemical identification and afferent

connections of subdivisions in the lateralis posterior-pulvinar

complex and related thalamic nuclei in the cat. Neuroscience

5:1175–1238.

Guildin WO, Markowitsch HJ (1984) Cortical and thalamic afferent

connections of the insular and adjacent cortex of the cat. J Comp

Neurol 229:393–418.

Guillery RW (1995) Anatomical evidence concerning the role of the

thalamus in cortico-cortical communication: a brief review. J Anat

187:583–592.

Hendry SHC, Jones EG, Graham J (1979) Thalamic relay nuclei for

cerebellar and certain related fiber systems in the cat. J Comp Neurol

185:679–714.

Henry GH, Salin PA, Bullier J (1991) Projections from areas 18 and 19 to

cat striate cortex: divergence and laminar specificity. Eur J Neurosci

3:186–200.

Hicks TP, Stark CA, Fletcher WA (1986) Origins of afferents to visual

suprageniculate nucleus of the cat. J Comp Neurol 246:544–554.

Higo S, Kawamura S (1984) Topographical linkage of tecto-

thalamo-anterior ectosylvian sulcal cortex in the cat: an 125I-WGA

autoradiographic study. Brain Res Bull 71:329–344.

Hilgetag CC, O’Neill MA, Young MP (1996) Indeterminate organization of

the visual-system. Science 271:776–777.

Hilgetag CC, Stephan KE, Burns GAPC, O’Neill MA, Young MP, Zilles K,

Kotter R (1997a) Systems of primate cortical areas defined by

functional connectivity. Soc Neurosci Abstr 514.13.

Hilgetag CC, O’Neill MA, Young MP (1997b) Optimization analysis of

complex neuroanatomical data. In: Computational neuroscience:

trends in research (Bower JM, ed.), pp. 925–930. New York: Plenum

Press.

Hilgetag CC, Burns GAPC, O’Neill MA, Young MP (1998) Cluster

structure of cortical systems in mammalian brains. In: Computational

neuroscience: trends in research (Bower JM, ed.). New York: Plenum

Press.

Hilgetag CC, Young MP, Scannell JW (1998) Models of paradoxical lesion

effects and rules for inference for imputing structure to function in

the brain. Proc. 7th Annual Computational Neuroscience Meeting.

Hirai T, Jones EG (1988) Segregation of lemniscal inputs and motor cortex

outputs in the cat ventral thalamic nuclei: application of a novel

technique. Exp Brain Res 71:329–344.

Hutchins B, Updyke BV (1988) The lateral posterior complex of the cat:

studies of the functional organization. Progr Brain Res 75:75–83.

Hutchins B, Updyke BV (1989) Retinotopic organization within the lateral

posterior complex of the cat. J Comp Neurol 285:350–398.

Imig TJ, Morel A (1983) Organization of the thalamocortical auditory

system in the cat. Annu Rev Neurosci 6:95–120.

Irle E, Markowitsch HJ (1982) Connections of the hippocampal-

formation, mamillary bodies, anterior thalamus and cingulate cortex

— a retrograde study using horseradish-peroxidase in the cat. Exp

Brain Res 47:79–94.

Jacobs RA, Jordan MI (1992) Computational consequences of a bias

towards short connections. J Cogn Neurosci 4:323–336.

Cerebral Cortex Apr/May 1999, V 9 N 3 295

Jimenez-Castellanos J, Reinoso-Suarez F (1985) Topographical organiza-

tion of the afferent connections of the principal ventromedial

thalamic nucleus in the cat. J Comp Neurol 236:297–314.

Johnstone T (1995) Magic 3D. London: Stanley Paul.

Jones EG (1985) The thalamus. New York: Plenum Press.

Jones EG, Burton H (1974) Cytoarchitecture and somatic sensory

connectivity of thalamic nuclei other than the ventrobasal complex in

the cat. J Comp Neurol 154:395–432.

Jones EG, Powell TPS (1968) The projection of the somatic sensory cortex

upon the thalamus in the cat. Brain Res 10:369–391.

Jones EG, Powell TPS (1969) The cortical projection of the

ventroposterior nucleus of the thalamus of the cat. Brain Res

13:298–318.

Jouve B, Rosenstiehl P, Imbert M (1998) A mathematical approach to the

connectivity between the cortical visual areas of the macaque

monkey. Cereb Cortex 8:28–39.

Kakei S, Shinoda Y (1990) Parietal projections of thalamocortical fibers

from the ventroanterior-ventrolateral complex of the cat thalamus.

Neurosci Lett 117:280–284.

Kalil RE, Tong L, Spear PD (1991) Thalamic projections to the lateral

suprasylvian visual areas in cats with neonatal or adult visual cortex

damage. J Comp Neurol 314:512–525.

Katayama T, Niimi K, Matsuoka H (1986) Thalamic projections to the

motor and premotor areas in the cat traced with horseradish

peroxidase. J Hirnforsch 27:45–59.

Kato N (1990) Cortico-thalamo-cortical projections between visual

cortices. Brain Res 509:45–59.

Kaufman EFS, Rosenquist AC (1985a) Efferent projections of the thalamic

intralaminar nuclei in the cat. Brain Res 335:257–279.

Kaufman EFS, Rosenquist AC (1985b) Afferent connections of the

thalamic intralaminar nuclei in the cat. Brain Res 335:281–296.

Kaufman EFS., Rosenquist AC, Rackowski D (1984) The projections of

single thalamic neurons onto multiple visual cortical areas in the cat.

Brain Res 298:171–174.

Kotter R (1992) The limbic system: a review of its empirical foundations.

Behav Brain Sci 52:105–127.

Kruskal JB (1964a) Multidimensional scaling by optimizing goodness of

fit to a nonmetric hypothesis. Psychometrika 19:1–27.

Kruskal JB (1964b) Non-metric multidimensional scaling: a numerical

method. Psychometrika 29:115–129.

Kuchiiwa W, Kuchiiwa T, Matusue H, Sukekawa K (1985) Efferent

connections of area 20 in the cat: HRP-WGA and autoradiographic

studies. Exp Brain Res 60:179–183.

Le Gros Clark WE (1932) The structure and connections of the thalamus.

Brain 55:261–273.

Le Gros Clark WE (1942) The visual centres of the brain and their

connections. Physiol Rev 22:205–232.

Lomber SG, Payne BR, Cornwell P (1996a) Learning and recall of form

discriminations during reversible cooling deactivation of

ventral-posterior suprasylvian cortex in the cat. Proc Natl Acad Sci

USA 93:1654–1658.

Lomber SG, Payne BR, Cornwell P, Long KD (1996b) Perceptual and

cognitive visual functions of parietal and temporal cortices in the cat.

Cereb Cortex 6:673–695.

Macchi G, Bentivoglio M, Minciacchi D, Molinari M, Norita M, Katoh Y

(1988) Synaptic organization of the lateralis-medialis-suprageniculate

nuclear (LM-Sg) complex in the cat. Progr Brain Res 75:09–119.

Macchi G, Bentivoglio M, Molinari M, Minciacchi D (1984) The

thalamo-caudate versus thalamo-cortical projections as studied in the

cat with f luorescent retrograde double labeling. Exp Brain Res

54:225–239.

MacNeil MA, Lomber SG, Payne BR (1997) Thalamic and cortical

projections to the middle suprasylvian cortex of cats: constancy and

variation. Exp Brain Res 114:24–32.

Martinez-Moreno E, Llamas A, Avendano C, Renes E, Reinoso-Suarez F

(1987) General plan of the thalamic projections to the prefrontal

cortex of the cat. Brain Res 407:17–26.

Masterton B (1976) Evolution of brain and behaviour in vertebrates. New

York: Lawrence Erlbaum.

Matsuoka H (1986) Topographic arrangement of the projections from the

anterior thalamic nuclei to the cingulate cortex in the cat. Neurosci

Res 4:62–66.

McCulloch WS (1944) The functional organization of the cerebral cortex.

Physiol Rev 24:390–407.

Mergner T, Deecke L, Wagner HJ (1981) Vestibulo-thalamic projection to

the anterior suprasylvian cortex of the cat. Exp Brain Res 44:455–458.

Meynert T (1890) Uber das Zusammenwirken der Gerhirntheile. In: Some

papers on the cerebral cortex (von Bonin G, ed.), pp. 173–190 [1960].

Springfield, IL: Thomas.

Micelli D, Reperant J, Marchand L, Ward R, Vesselkin N (1991)

Divergence and collateral axon branching in subsystems of visual

cortical projections from the cat lateral posterior nucleus. J Hirnforsch

32:165–173.

Middlebrooks JC, Zook JM (1983) Intrinsic organization of the cat’s

medial geniculate body defined by projections to binaural

response-specific bands in the primary auditory cortex. J Neurosci

3:203–224.

Minciacchi D, Bentivoglio M, Ilinsky IA, Kultasilinsky K, Macchi G,

Molinari M (1986) Multiple cortical targets of one thalamic nucleus —

the projections of the ventral medial nucleus in the cat studies with

retrograde tracers. J Comp Neurol 252:106–129.

Mitani A, Itoh K, Kaneko T, Kudo M, Mizuno N, Nomura S (1984)

Thalamocortical projections to layer I of the primary auditory cortex

in the cat: a horseradish-peroxidase study. Brain Res 310:347–350.

Mitani A, Itoh K, Mizuno N (1987) Distribution and size of thalamic

neurons projecting to layer I of the auditory cortical fields of the cat

compared to those projecting to layer IV. J Comp Neurol

257:105–121.

Mitchison G (1991) Neuronal branching and the economy of cortical

wiring. Proc R Soc Lond B 245:151:158.

Miyashita E, Tamai Y (1989) Subcortical connections of frontal

‘occulomotor’ areas in the cat. Brain Res 502:755–787.

Morel A, Imig TJ (1987) Thalamic projections to fields A, AI, P, VP in the

cat auditory cortex. J Comp Neurol 265:119–144.

Movshon JA, Adelson EH, Gizzi MS, Newsome WT (1985) The analysis of

moving visual patterns. In: Experimental Brain Research, suppl. vol.

11: Pattern recognition mechanisms (Chagas C, Gattas R, Gross C,

eds), pp. 117–151.

Musil SY, Olson CR (1988a) Organization of the cortical and subcortical

projections to the medial prefrontal cortex in the cat. J Comp Neurol

272:219–241.

Musil SY, Olson CR (1988b) Organization of the cortical and subcortical

projections to the anterior cingulate cortex in the cat. J Comp Neurol

272:203–218.

Musil SY, Olson CR (1991) Cortical areas in the medial frontal-lobe of the

cat delineated by quantitative-analysis of thalamic afferents. J Comp

Neurol 308:457–466.

Musil SY, Olson CR (1993) The role of cat cingulate cortex in

sensorymotor integration. In: Neurobiology of cingulate cortex and

limbic thalamus: a comprehensive handbook (Vogt BA, Gabriel M,

eds), pp. 345–365. Boston, MA: Birkhauser.

Naito J, Kawamura K (1982) Thalamocortical neurons projecting to the

areas surrounding the anterior and middle suprasylvian sulci in the

cat. Exp Brain Res 45:59–70.

Nicolelis MAL, Tinone G, Sameshima K, Timo-Iaria C, Hong YC, Van de

Bilt MT (1990) Connection, a microcomputer program for storing and

analyzing structural properties of neural circuits. Comput Biomed Res

23:64–81.

Niimi M (1978) Cortical projections of the anterior thalamic nuclei in the

cat. Exp Brain Res 31:403–416.

Niimi K, Niimi M, Okada Y (1978) Thalamic afferents to the limbic cortex

in the cat studies with the method of retrograde ozonal transprt of

horseradish peroxidase. Brain Res 145:225–238.

Niimi K, Imataki T, Ono K, Matsuoka H, Miyata Y (1987) Thalamic

projections to the second and fourth somesthtic areas in the anterior

ectosylvian gyrus of the cat. J Hirnforsch 28:271–280.

Niimi K, Miyata Y, Matsuoka H (1989) Thalamic projections to the cortical

gustatory area in the cat studies by retrograde axonal transport of

horseradish peroxidase. J Hirnforsch 30:589–593.

Norita M, Mucke L, Benedek G, Albowitz B, Katoh Y, Creutzfeldt OD

(1986) Connections of the anterior ectosylvian visual area. Exp Brain

Res 62:225–240.

Ojima H (1994) Terminal morphology and distribution of corticothalamic

fibers originating from layer-5 and layer-6 of the cat primary auditory

cortex. Cereb Cortex 4:646–663.

O’Leary DDM, Schlaggar BL, Tuttle R (1994) Specification of neocortical

areas and thalamocortical connections. Annu Rev Neurosci

17:419–439.

Olson CR, Lawler K (1987) Cortical and subcortical afferent connections

296 Cat Cortico-thalamic Network • Scannell et al.

of a posterior division of feline area 7 (area 7p). J Comp Neurol

259:13–30.

Olson CR, Jeffers IM (1988) Organization of cortical and subcortical

projections to area 6m in the cat. J Comp Neurol 266:73–94.

Olson CR, Musil SY (1992) Topographic organization of cortical and

subcortical projections to posterior cingulate cortex in the cat —

evidence for somatic, ocular, and complex subregions. J Comp Neurol

324:237–260.

Ono K, Niimi K (1986) Afferent projections to the thalamic mediodorsal

nucleus in the cat studied by retrograde and anterograde axonal

transport of horseradish peroxidase. J Hirnforsch 27:597–610.

Pandya DN, Yeterian EH (1985) Architecture and connections of cortical

association areas. In: Cerebral cortex, vol. 4 (Peters A, Jones EG, eds),

pp. 3–61. New York: Plenum Press.

Payne BR, Pearson HE, Cornwell P (1988) Development of visual and

auditory cortical connections in the cat. In: Cerebral cortex, vol. 7:

Development and maturation of the cerebral cortex (Peters A, Jones

EG, eds), pp. 309–389. New York: Plenum Press.

Penny GR, Diamond IT, Itoh K (1982) Cells of different sizes in the ventral

nuclei project to different layers of the somatic cortex in the cat. Brain

Res 242:55–65.

Peters A, Payne BR (1993) Numerical relationships between geniculo-

cortical afferents and pyramidal cell modules in cat primary

visual-cortex. Cereb Cortex 3:69–78.

Peters A, Payne BR, Budd J (1994) A numerical-analysis of the

geniculocortical input to striate cortex in the monkey. Cereb Cortex

4:215–229.

Price DJ, Zumbroich TJ (1989) Postnatal development of cortico-cortical

efferents from area 17 in the cats’ visual cortex. J Neurosci 9:600–613.

Raczkowski D, Rosenquist AC (1980) Connections of the parvocellular

C-laminae of the dorsal lateral geniculate nucleus with the visual

cortex in the cat. Brain Res 199:447–451.

Raczkowski D, Rosenquist AC (1983) Connections of the multiple visual

cortical areas with the lateral posterior-pulvinar complex and adjacent

thalamic nuclei in the cat. J Neurosci 3:1912–1942.

Rajkowska G, Goldman-Rakic PS (1995) Cytoarchitectonic definition of

prefrontal areas in the normal human cortex. 2. Variability in the

locations of areas 9 and 46 and relationships to the Talairach

coordinate system. Cereb Cortex 5:323–337.

Ramirez-Camacho R, Avendano C, Reinoso-Suarez F (1984) Thalamic

projections to the anterior suprasylvian and posterior sigmoid cortex:

an HRP study of the ‘vestibular areas’ of the cerebral cortex in the cat.

Brain Res Bull 12:245–252.

Ramon-Moliner E (1984) Subcortical projections of the pericruciate

cortex of the cat. An autoradiographic study. J Hirnforsch

25:445–459.

Reinoso-Suarez F (1984) Connections patterns in parieto-temporo-

occipital association cortex of the feline cerebral cortex. In: Cortical

integration (Reinoso-Suarez F, Ajmone-Mersan C, eds), pp. 255–278.

New York: Raven Press.

Robertson RT, Cunningham TJ (1981) Organization of corticothalamic

projections from parietal cortex in the cat. J Comp Neurol

199:569–585.

Rockland KS, Pandya DN (1979) Laminar origins and terminations of

cortical connections of the occipital lobe in the rhesus monkey. Brain

Res 179:3–20.

Roda JM, Reinoso-Suarez F (1983) Topographical organization of the

thalamic projections to the cortex of the anterior ectosylvian sulcus in

the cat. Exp Brain Res 49:131–139.

Rodrigues-Dagaeff C, Simm G, De Ribaupierre Y, Villa A, De Ribaupierre

F, Rouiller EM (1989) Functional organization of the medial division of

the medial geniculate body of the cat: rvidence for a rostro-caudal

gradient of response properties and cortical projections. Hear Res

39:103–126.

Room P, Groenewegen HJ (1986) Connections of the parahippocampal

cortex in the cat. II. Subcortical afferents. J Comp Neurol 251:

451–473.

Room P, Russhen FT, Groenewegen HJ, Lohman AHM (1985) Efferent

connections of the prelimbic (area 32) and the infralimbic (area 25)

cortices: an anterograde tracing study in the cat. J Comp Neurol

242:40–55.

Rosenquist AC (1985) Connections of visual cortical areas in the cat. In:

Cerebral cortex, vol. 3: Visual cortex (Peters A, Jones EG, ed), pp.

81–117. New York: London.

Rouiller EM, de Ribaupierre Y (1985) Origin of afferents to physio-

logically defined regions of the medial geniculate body of the cat —

ventral and dorsal divisions. Hear Res 19:97–114.

Rouiller EM, de Ribaupierre F (1990) Arborization of corticothalamic

axons in the auditory thalamus of the cat: a PHA-L tracing study.

Neurosci Lett 108:29–35.

Rouiller EM, Rodrigues-Dagaeff C, Simm G, de Ribaupierre Y, Villa, A, de

Ribaupierre F (1989) Functional organization of the medial division of

the medial geniculate body of the cat: tonotopic organization, spatial

distribution of response properties and cortical connections. Hear Res

39:127–142.

Royce GJ (1983) Cortical neurons with collateral projectiosn to both the

caudate nucleus and centromedian-parafasicular thalamic complex: a

f luorescent retrograde double labeling study in the cat. Exp Brain Res

50:157–165.

Royce GJ, Mourey RJ (1985) Efferent connections of the centromedial and

parafasicular thalamic nuclei: an autoradiographic investigation in the

cat. J Comp Neurol 235:2770–300.

Royce GJ, Bromley S, Gracco C, Beckstead RM (1989) Thalamocortical

connections of the rostral intralaminar nuclei: an autoradiographic

analysis in the cat. J Comp Neurol 288:555–582.

Savage RJG (1986) Mammalian evolution: an illustrated guide. London:

British Museum (Natural History).

Scannell JW (1996) Why sulci and gyri are where they are. Brain Res

Assoc Abstr 13:24.07.

Scannell JW (1997) Determining cortical landscapes. Nature 386:452.

Scannell JW, Young MP (1993) The connectional organization of neural

systems in the cat cerebral-cortex. Curr Biol 3:191–200.

Scannell JW, Blakemore C, Young MP (1995) Analysis of connectivity in

the cat cerebral-cortex. J Neurosci 15:1463–1483.

Scannell JW, Sengpiel F, Tovee MJ., Benson PJ, Blakemore C, Young MP

(1996) Visual-motion processing in the anterior ectosylvian sulcus of

the cat. J Neurophysiol 76:895–907.

Scannell JW, Burns G, O’Neill MA, Hilgetag CC, Young MP (1997) The

organization of the thalamo-cortical network of the cat. Soc Neurosci

Abstr 514.12.

Schonemann P, Carrol RM (1970) Fitting one matrix to another under a

choice of a similarity transformation and rigid motion. Psychometrika

35:245–255.

Sherk H (1986) Location and connections of visual cortical areas in the

cats’ suprasylvian sulcus. J Comp Neurol 247:1–31.

Shinoda Y, Kakei S (1989) Distribution of terminals of thalamocortical

fibers originating from the ventrolateral nucleus of the cat thalamus.

Neurosci Lett 96:163–167.

Shinoda Y, Kakei S, Futami T, Wannier T (1993) Thalamocortical

organization in the cerebello-thalamo-cortical system. Cereb Cortex

3:421–429.

Shinonaga Y, Takada M, Mizuno N (1994) Direct projections from the

non-laminated divisions of the medial geniculate-nucleus to the

temporal polar cortex and amygdala in the cat. J Comp Neurol

340:405–426.

Silvermann BW (1986) Density estimation for statistics and data analysis.

London: Chapman and Hall.

Simmen MW, Goodhill GJ, Willshaw DJ (1994) Scaling and brain

connectivity. Nature 369:448–449.

Spreafico R, Hayes NL, Rustioni A (1981) Thalamic projections to the

primary and secondary somatosensory cortices in the cat: single and

double retrograde tracer studies. J Comp Neurol 203:67–90.

Stein BE, Edwards SB, Spencer RF (1983) Corticotectal and cortico-

thalamic efferent projections of SIV somatosensory cortex in the cat.

J Neurophysiol 50:896–909.

Symonds LL, Rosenquist AC (1984) Laminar origins of visual cortico-

cortical connections in the cat. J Comp Neurol 229:39–47.

Symonds LL, Rosenquist AC, Edwards SB, Palmer LA (1981) Projections of

the pulvinar-lateral posterior complex to visual cortical areas in the

cat. Neuroscience 6:1995–2020.

Takane Y, Young FW, de Leeuw J (1977) Non-metric multidimensional

scaling: an alternating least squares method with optimal scaling

features. Psychometrika 42:7–67.

Tamai Y, Miyashita E (1989) Subcortical connections of a ‘oculomotor’

region in the ventral bank of the anterior ectosylvian sulcus in the cat.

Neurosci Res 7:249–256.

Tanibuchi I (1992) Electrophysiological and anatomical studies on

thalamic mediodorsal nucleus projections onto the prefrontal cortex

in the cat. Brain Res 580:137–146.

Tanji DG, Wise SP, Dykes RW, Jones EG (1978) Cytoarchitecture and

Cerebral Cortex Apr/May 1999, V 9 N 3 297

thalamic connectivity of third somatosensory area of the cat cerebral

cortex. J Neurophysiol 41:268–284.

Tong L, Kalil RE, Spear PD (1982) Thalamic projections to visual areas of

the middle suprasylvian sulcus in the cat. J Comp Neurol

212:103–117.

Ungerleider LG, Mishkin M (1982) Two cortical visual systems. In:

Analysis of visual behavior (Ingle DG, Goodale MA, Mansfield RJQ,

eds), pp 549–586. Boston, MA: MIT Press.

Updyke BV (1981) Projections from visual areas of the middle

suprasylvian sulcus onto the lateral posterior complex and adjacent

thalamic nuclei in the cat. J Comp Neurol 201:477–506.

Updyke BV (1983) A re-evaluation of the functional organization and

cytoarchitecture of the feline lateral posterior complex, with

observations on adjoining cell groups. J Comp Neurol 219:143–181.

Van Essen DC (1997). A tension based theory of morphogenesis and

compact wiring in the central nervous system. Nature 385, 313–318.

VanDuffel W, Payne BR, Lomber SG, Orban G (1997) Functional impact of

cerebral connections. Proc Acad Sci USA 94:7617–7620.

Van Laarhoven PJM, Aarts HL (1987) Simulated annealing: theory and

applications. Dordrecht: Kluwer.

von Bonin G B. P (1947) The neocortex of Macaca mulatta. Urbana:

Univeristy of Illinois Press.

Wannier T, Kakei S, Shinoda Y (1992) Two modes of cerebellar input to

the parietal cortex in the cat. Exp Brain Res 90:241–252.

Webster MJ, Bachevalier J, Ungerleider LG (1994) Connections of inferior

temporal areas TEO and TE with parietal and frontal cortex in

macaque monkeys. Cereb Cortex 5:470–483.

Winer JA (1992) The functional architecture of the medial geniculate

body and the primary auditory cortex. In: The mammalian auditory

pathway: neuroanatomy (Webster DB, Popper AN, Fay RR, eds), pp.

222–409. New York: Springer Verlag.

Winer JA, Morest DK (1983a) The neuronal architecture of the dorsal

division of the medial geniculate body of the cat. A study with the

rapid golgi method. J Comp. Neurol 221:1–30.

Winer JA, Morest DK (1983b) The medial division of the medial geniculate

body of the cat: implications for thalamic organization. J Neurosci

3:2629–2651.

Winer JA, Diamond IT, Raczkowski D (1977) Subdivisions of the auditory

cortex of the cat: the retrograde transport of horseradish peroxidase

to the medial genicularte body and posterior thalamic nuclei. J Comp

Neurol 176:387–418.

Witter MP, Groenewegen HJ (1986) Connections of the parahippocampal

cortex in the cat. II. Cortical and thalamic efferents. J Comp Neurol

252:1–31.

Yamamoto T, Kishimoto Y, Takagi H, Yoshikawa H, Oka H (1992)

Morphological features of cat thalamo-parietal fibres investigated by

PHA-L immunohistochemistry and electron microscopy. Neurosci Res

14:124–132.

Yanagihara M, Niimi K, Ono K (1987) Thalamic projections to the

hippocampal and entorhinal areas in the cat. J Comp Neurol

266:122–141.

Yasui Y, Itoh K, Kaneko T, Sugimoto T, Mizuno N (1986) Direct

projections from the parvocellular part of the posteromedial ventral

nucleus of the thalamus to the infralimbic cortex in the cat. Brain Res

368:384–388.

Yasui Y, Itoh K, Kaneko T, Mizuno N, Sugimoto T (1987) Thalamocortical

and thalamo-amygdaloid projections from the parvicellular division of

the posteromedial ventral nucleus in the cat. J Comp Neurol

257:253–268.

Yasui Y, Itoh K, Kamiya H, Ino, T, Mizuno N (1988) Cingulate gyrus of the

cat recieves projection fibers from the thalamic region ventral to the

ventral border of the ventrobasal complex. J Comp Neurol

274:91–100.

Young FW, Harris DF (1990) SPSS base sytem user’s guide (Norusis MJ,

ed.). Chicago, IL: SPSS Inc.

Young FW, Takane Y, Lewyckyj R (1978) ALSCAL: a nonmetric

multidimensional scaling program with several differences opotions.

Behav Res Methods Instrum 10:451–453.

Young MP (1992) Objective analysis of the topological organization of the

primate cortical visual-system. Nature 358:152–155.

Young MP (1993) The organization of neural systems in the primate

cerebral-cortex. Proc R Soc Lond B Biol Sci 252:13.

Young MP, Scannell JW, Burns GAPC, Blakemore C (1994) Scaling and

brain connectivity — reply. Nature 369:449–450.

Young MP, Scannell JW, Burns APC, Blakemore C (1995a) Analysis of

connectivity: systems in the cerebral cortex. Rev Neurosci 227–250.

Young MP, Scannell JW, Burns GAPC (1995b) The analysis of cortical

connectivity. Austin, TX: R.G. Landes.

Young MP, Scannell JW, O’Neill MA, Hilgetag CC, Burns G, Blakemore C

(1995c) Nonmetric multidimensional-scaling in the analysis of

neuroanatomical connection data and the organization of the primate

cortical visual-system. Phil Trans R Soc Lond B Biol Sci 348:281–308.

Zheng ZH, Wu CP, Xi MC (1986) Thalamic projections to motor area and

area 3a of the cat cerebral cortex. Brain Res 380:389–393.

AppendixParcellation of the cat thalamus

Thalamic division Be&Jo 1982 Abbr. A-P M-L D-V Notes

Medial geniculate complexVentral division of the MG MGP MGV 5.4 10.0 0.5 equivalent to LV + OV (Mo64;Ca83; Ca&Ai83), Vl + Vo (An80), V (Mo&Im57), VD (Wi&Mo83a),

CGMv (Wi&Mo83b; Wi92); also includes ‘d’: dorsal cap nucleus of Mo&Im87Dorsal division of the MG MGD 4.5 9.0 2.0 equivalent to CGMd (Wi&Mo83b; Wi92)Deep dorsal nucleus MGDdd 4.5 9.0 1.0 equivalent to DP (Mo64), Dd (An80; Ca83; Calford and Aitkin, 1983; Wi92), DD (Wi&Mo83a,b;

Wi92)Dorsal nucleus MGDd 4.5 9.0 2.0 equivalent to D (Mo64; Wi&Mo83a,b; Wi92) and part of Dc (An80) and DC (Ca83; Ca&Ai83);

includes ‘cd’: caudodorsl nuleus of Mo&Im87Superficial dorsal nucleus MGDds 4.5 9.5 2.2 equivalent to DS (Mo64; Wi&Mo83a,b; Wi92), part of Dc (An80) and DC (Ca83; Ca&Ai83)Ventrolateral nucleus MGvl 4.8 9.5 –1.0 eqivalent to VL (Mo64; An80;Wi&Mo83a,b; Ca83; Ca&Ai83; Wi92), vl (Mo&Im87)

Magnocellular medial geniculate nucleus MGM MGM 6.4 7.5 2.0

Dorsal lateral geniculate complex LGLaminar dorsal lateral geniculate nucleus LGLA laminae LGLA 6.0 9.5 6.0 includes laminae A and A1 (e.g. Dr86; Ro85)C lamina LGLC 6.0 9.5 5.3 the Magno C laminae (Ro85; Dr86)C1-3 laminae LGLC1–3 6.0 9.5 4.7 The Parvo C laminae (Ro85; Dr86)Geniculate wing LGLGW 7.2 7.5 7.0 an extension of the Parvo C laminae onto the surface of the Pulvinar (Ra83; Gu80)Medial interlaminar nucleus LGI LGLI 6.4 7.7 5.2 equivalent to MIN (Ro85; Dr86)

Posterior complex PoMedial region POM POm 8.0 5.5 4.0 equivalent to POm of Ba91 and Bu&Ko84b

Lateral region POL POl 8.3 8.0 3.0 equivalent to Po of Mo&Im87 and Po of Wi92Suprageniculate nucleus SGN SG 6.4 5.5 3.5 equivalent to NS (Ca83; Ca&Ai83); partially overlaps with LM-SgIntermediate region POI POi 6.0 8.0 3.5 equivalent to PoI (Cl&Ir90; Ro&Re83) and POI (Gu&Ma84)

298 Cat Cortico-thalamic Network • Scannell et al.

Appendix – continued

Thalamic divsion Be&Jo 1982 Abbr. A-P M-L D-V Notes

Lateral nucleiLateral dorsal nuclei LD LD 9.5 3.5 9.3 equivalent to LD (Av90; Av85)Lateral posterior nucleus LPLateral division of LP LPl 6.8 5.0 8.0 equivalent to the ‘corticorecipient zone’ of Dr85 and LPl and PNR of Gr&Be80; corresponds to

postero-lateral part of LPl of Up83Medial division of LP LPm 6.8 4.5 7.5 equivalent to the ‘tectorecipient zone’ of Dr85, to LPm of Gr&Be80 and to LPi of Up83Lateral medial and suprageniculate complex LM-Sg 6.4 4.5 5.0Pulvinar PUL PUL 7.2 6.5 8.0 equivalent to ‘pretectorecipient zone’ of Dr85 and the pulvinar (Up83; Gr80)Posterior nucleus of Rioch PNR 7.2 7.5 5.5 NP of Gr&Be80 is connectionally similar to LPl, with which it has been lumped in this studyIntermediate nucleus of the lateral group LI LI of Gr&Be80 corresponds to the rostral regions of LPl or Up83Caudal part of LI LIc 8.3 4.0 8.0 LIc (Gr&Be80; Av85); rostral part of LPl of Up83Oral part of LI LIo 9.5 5.5 7.0 LIo (Gr&Be80; Av85); rostral part of LPl of Up83Mediodorsal nucleus MD MD 9.1 2.0 4.5

Ventroposterior complex VBLateral part of VP (spinothalamic) VBX VPL 9.1 7.0 3.5 equivalent to VPL (Bu&Ko84b; Zh86; Ba91); includes VPL of Dy86Medial part of VP(trigeminal) VBA VPM 9.1 5.5 4.0 equivalent to VPM (Bu&Ko84b; Zh86; Ba91); includes VPM of Dy86External part of VP VPX

VPo VPO of Dy86: fine-grained division that has not been demonstrated connectionally in other studiesVPI VPI of Dy86: fine-grained division that has not been demonstrated connectionally in other studiesVPS VPS of Dy86: fine-grained division that has not been demonstrated connectionally in other studies

Ventroanterior-ventrolateral nucleus VA&VL VA-VL 11.8 5.5 5.0Ventromedial complex

Principal ventromedial nucleus VMP VMP 10.5 2.0 3.0Basal ventromedial nucleus VMB VMB 8.3 3.5 2.0 equivalent to nucleus ventralis medialis pars parvocellularis (e.g. Ru72)Submedial nucleus SUM SUM 10.6 1.0 3.5

Rostral intralaminar nucleiCentral medial nucleus CMN CMN 10.2 0.0 4.0Paracentral nucleus PAC PAC 9.5 2.0 4.1Central lateral nucleus CLN CLN 10.2 3.5 6.0

Caudal intralaminar nucleiCentre median nucleus CM CM 7.2 2.7 4.2Parafasicular nucleus PF PF 8.0 1.5 3.5Rhomboid nucleus RH 11.0 0.0 4.8Parataenial nucleus PAT 12.0 1.0 5.5Anterior paraventricular nucleus PARA 11.8 0.5 6.0Posterior paraventricular nucleus PARP 9.5 0.5 5.8Medioventral-reuniens nucleus MV MV-RE 12.0 0.0 2.0 known alternatively as the reuniens or the medioventral nucleus

Anterior complexAnterodorsal nucleus AD AD 12.0 2.0 6.0Anteroventral nucleus AV AV 12.0 3.0 6.5Anteromedial nucleus AM AM 12.0 1.3 3.8

The first column of the table shows the major thalamic divisions and subdivisions. The second column (Be&Jo, 1982) gives the nomenclature used in the Berman and Jones (1982) atlas. The third columnshows the nomenclature used in this paper. Where possible, we have used the same names as Berman and Jones (1982). Nuclei shown in bold typeface in the ‘Abbr.’ column had sufficient connection datato be included in the quantitative analyses. To help identify our parcellation and its relation to other schemes, the stereotaxic coordinates of the approximate centres of the nuclei, taken from the frontalsections of Berman and Jones (1982), are also given. ‘A-P’ corresponds to anteroposterior coordinate, ‘M-L’ to mediolateral coordinate, and ‘D-V’ to dorsoventral coordinate. The final column of the tableprovides notes on alternative nomenclature and parcellation. The references are as follows: An80 = Andersen et al. (1980); Av85 = Avendano et al. (1985); Av90 = Avendano et al. (1990); Ba81 =Barbaresi et al. (1981); Bu&Ko84b = Burton and Kopf (1984b); Ca&Ai83 = Calford and Aitkin (1983); Ca83 = Calford (1983); Cl&Ir90 = Clarey and Irvine (1990); Dr86 = Dreher (1986); Dy86 = Dykes etal. (1986); Gr&Be80 = Graybiel and Berson (1980); Gu&Ma84 = Guildin and Markowitsch (1984); Gu80 = Guillery et al. (1980); Mo&Im87 = Morel and Imig (1987); Mo64 = Morest (1964); Raczkowskiand Rosenquist (1983); Ro&Re83 = Roda and Reinoso-Suarez (1983); Ro85 = Rosenquist (1985); Ru72 = Ruderman et al. (1972); Up83 = Updyke (1983); Wi&Mo83a = Winer and Morest (1983a);Wi&Mo83b = Winer and Morest (1983b); Wi91 = Winer (1991); Zh86 = Zheng et al. (1986).

Cerebral Cortex Apr/May 1999, V 9 N 3 299