Image Browsing in Slow Medical Liquid Crystal Displays

Technical Report

Image Browsing in Slow MedicalLiquid Crystal Displays1

Hongye Liang, PhD, Subok Park, PhD, Brandon D. Gallas, PhD, Kyle J. Myers, PhD, Aldo Badano, PhD

Rationale and Objectives. Statistics show that radiologists are reading more studies than ever before, creating the chal-lenge of interpreting an increasing number of images without compromising diagnostic performance. Stack-mode imagedisplay has the potential to allow radiologists to browse large three-dimensional (3D) datasets at refresh rates as high as30 images/second. In this framework, the slow temporal response of liquid crystal displays (LCDs) can compromise theimage quality when the images are browsed in a fast sequence.

Materials and Methods. In this article, we report on the effect of the LCD response time at different image browsingspeeds based on the performance of a contrast-sensitive channelized-hoteling observer. A stack of simulated 3D clusteredlumpy background images with a designer nodule to be detected is used. The effect of different browsing speeds is calcu-lated with LCD temporal response measurements from our previous work. The image set is then analyzed by the modelobserver, which has been shown to predict human detection performance in Gaussian and non-Gaussian lumpy back-grounds. This methodology allows us to quantify the effect of slow temporal response of medical liquid crystal displayson the performance of the anthropomorphic observers.

Results. We find that the slow temporal response of the display device greatly affects lesion contrast and observer perfor-mance. A detectability decrease of more than 40% could be caused by the slow response of the display.

Conclusions. After validation with human observers, this methodology can be applied to more realistic background datawith the goal of providing recommendations for the browsing speed of large volumetric image datasets (from computedtomography, magnetic resonance, or tomosynthesis) when read in stack-mode.

Key Words. Model observer; display quality; contrast sensitivity; temporal response.©
AUR, 2008
The high-fidelity display of medical images helps radiolo-gists make correct diagnoses. Active-matrix liquid crystaldisplays (LCDs) are taking the place of traditional cath-

Acad Radiol 2008; 15:370–382

1 CDRH/NIBIB Laboratory for the Assessment of Medical Imaging Systems,Division of Imaging and Applied Mathematics, Office of Science and Engi-neering Laboratories, Center for Devices and Radiological Health, U.S.Food and Drug Administration, 10903 New Hampshire Avenue, Building 62,Room 3113, Silver Spring, MD 20993 (H.L., S.P., B.D.G., K.J.M., A.B.). Re-ceived September 9, 2007; accepted October 20, 2007. The mention ofcommercial products herein is not to be construed as either an actual orimplied endorsement of such products by the Department of Health andHuman Services. This is a contribution of the US Food and Drug Adminis-tration, and is not subject to copyright. Address correspondence to: A.B.e-mail: [email protected]

©
AUR, 2008doi:10.1016/j.acra.2007.10.017
370

ode-ray tube (CRT) displays and hard-copy films in mostmedical diagnostic imaging modalities. The image qualityof LCD monitors has been significantly improved overthe last few years. High-performance LCDs are consid-ered to have comparable or better performance than CRTsin displaying static images for all aspects of displayperformances except noise and viewing angle (1,2).However, LCDs can be inferior to CRTs at displayingmoving scenes due to their slow response and the hold-type addressing scheme used in their pixel addressingmethods (3,4).

With the increasing number of computed tomography(CT) images to be interpreted per day, the ability ofshowing a fast sequence of images in stack-mode has
been described as a preferred choice for a medical dis-
mailto:[email protected]

Academic Radiology, Vol 15, No 3, March 2008 IMAGE BROWSING IN SLOW MEDICAL LIQUID CRYSTAL DISPLAYS

play. When using stack-mode browsing, slices are stackedon top of each other in the display memory buffer. At agiven point, only one image is visible on the screen withthe images being browsed back and forth at high imagerefresh rates under the control of the radiologist. When ahigh browsing speed is used, the radiologist perceives athree-dimensional (3D) volumetric image instead of a se-ries of separate slices. The refresh rate could reach 30images per second (ie, the duration for each specific im-age on the screen is only 33 milliseconds) depending onthe application and equipment used (5,6). Further increaseof browsing speed with the advancement of technology isexpected.

For most medical displays, temporal response is notyet a topic of consideration. However, the slowest inter–gray-level response time of medical LCDs in use canreach more than 140 milliseconds, with most of the tran-sitions lying in the region of 40–100 milliseconds. Whenthe refresh rate of stack-mode image reading reaches 30images/second or higher, the LCD response time ofaround 60–100 milliseconds is, as we show in this article,far from ideal. In this article, we investigate the effect ofslow display response when using reasonable refresh ratesfor medical displays with different temporal performance.

The temporal response of LCDs depends on the liquidcrystal (LC) cell gap width, on the driving voltage to theLC cell, and on the LC material properties (ie, rotationalviscosity, dielectric anisotropy, and elastic constants) (7).Higher voltage or narrower LC cell gap leads to higherelectric fields causing the LC molecules to move faster.The mechanical analogy is pulling a spring hard whichmakes it retract faster when released. This explains thefact that black-to-white transition times are usually lessthan inter–gray-level transition times for LCDs. LC mate-rial viscosity and a higher elastic constant can also limitthe response time. The intrinsic inter–gray-level responsetime is typically on the order of several tens to hundredsof milliseconds. Different techniques can reduce the re-sponse time. A direct way of improving the LCD tempo-ral performance is by using faster materials and smallerLC gaps, but this could be very costly. A more popularalternative adopted by display manufacturers to improveLCD temporal performance is to use modified addressingschemes including overdrive techniques (8). These tech-niques are based on the fact that larger pixel voltages leadto faster transitions between gray levels. In general, theluminance level of a specific gray level cannot beachieved within one frame time because of the slow re-
sponse of LC molecules. If a higher voltage is applied to
the pixel during a fraction of the frame time, the LC mol-ecules can move faster, and thus the target luminancelevel might be achieved within one frame time. However,this technique is not yet used in medical displays becauseof implementation complexity.

The slow temporal response of a display can affect theexact mapping of the image pixel values into luminancevalues. For a slow display, the targeted pixel luminanceoften cannot be achieved within a frame time. Figure 1shows possible luminance transitions for pixel (i,j). Thepixel has a luminance of L(n � 1) at frame n � 1. Atframe n, the pixel luminance is set to change to L0(n).Because of the slow response time, it takes the displayseveral frames to reach the target luminance (indicated by 0).Thus, at the end of frame n, the pixel only reaches L(n).When the pixel value changes for every frame, it is possiblethat it will never reach the intended luminance. Luminancetransition profiles may take different forms as indicated bythe dashed, dotted, and dash-dotted lines in Fig 1. Eachcurve depicts a different response time profile, but theycould all lead to the same luminance at the frame end,which makes the analysis of temporal performance of LCDseven more complicated.

In this article, we describe a method for quantifyingthe effect of slow response of medical LCDs based on theperformance of an anthropomorphic model observer for

Figure 1. Luminance change for a dynamic scene. L0(i, j, n) indi-cates the target luminance of pixel (i, j) at frame n, correspondingto a gray level g0(i, j, n). At frame n � 1, the target gray level isg0(i, j, n � 1) (target luminance level L0(i, j, n � 1)). Because ofslow temporal response, the achieved luminance at frame n � 1is L(i, j, n � 1). The profile of luminance change from frame n to n� 1 may take different forms, as indicated by the dashed lines,depending on different liquid crystal display technologies and dif-ferent gray level transitions.

detection of a nodule in a complex background. We relied

371

LIANG ET AL Academic Radiology, Vol 15, No 3, March 2008

on a mathematical model observer that predicts humanperformance. Mathematical observers, ideal (9) and an-thropomorphic (10), are used to guide the development ofmedical imaging systems. The use of models for the hu-man vision system alleviates the need to conduct demand-ing reader studies that most often are plagued with largereader variability. In this article, we provide a first at-tempt to determine the effect of slow temporal responseof medical displays on detection tasks. For this work, theobserver of choice is an anthropomorphic model for thedetection task based on channelized-hoteling observer(CHO) for which we and others have developed signifi-cant experience in a variety of visual tasks in medicalimaging (11–13). Although the method has been previ-ously validated for similar signal tasks, more researchwill be completed in a follow-up article to validate theresults presented here. In particular, in the Discussionsection, we declare the limitations of our current approachand suggest several directions for improvement of themethod with respect to make it more applicable to realisticimaging data. However, the data presented in this articlerepresent an initial useful description of the potential prob-lems associated with slow displays and should be considereda starting point for more elaborate approaches that will beneeded to develop a complete understanding of the limita-tions of stack-mode reading of large image datasets.

MATERIALS AND METHODS

MeasurementsTypically, LCDs have gray level transition times of

several tens of milliseconds depending on the specificgray level pair. We start by measuring gray level transi-tion times for all possible transitions (for an eight-bit pal-ette) to fully characterize the LCDs tested in this article.

The measurement is based on a gray level transitiontime measurement system (Westar Display Technologies,St. Charles, MO) composed of a display driver (T-drive)for display pattern generation and a photomultiplier tube(PMT, model TRD-100A) for light detection (14). Allmeasurements are carried out in a dark room to avoidsignal from ambient light sources. The LCD is driven tochange its emission between two gray levels. At eitherlevel, the LCD is driven long enough (0.5 seconds) tocomplete the luminance transition. The luminance transi-tion waveform is recorded by the PMT. By changing thegray levels from 0 to 255, we can measure transition
times between any gray level pair in a 256 � 256 matrix.
372

At the same time, the measurement system is set torecord the percentage of luminance achieved (ratio of lu-minance at the end of the frame to the target luminance).A “percentage matrix” (P(g1, g2)) reflecting the actualpercentage of luminance reached after one frame is gener-ated and used for the simulations. Obtaining this responsetime matrix generally takes more than 20 hours of contin-uous measurements. Details on the measurements can befound elsewhere (14).

Simulation of Temporal EffectsThe images displayed on LCDs with slow response

can be modeled using the measured percentage matrix.When images are displayed on a slow LCD at a high re-fresh rate, the image quality will degrade in the form ofreduced luminance contrast. Our method simulates imagesbeing displayed on an LCD at different browsing speeds,resulting in different degrees of degradation. The simu-lated images can be used to assess the display temporalperformance and to determine a proper browsing speedfor accurate stack-mode medical image reading.

A choice of structured background for studying theeffect of slow display on stack-mode reading needs toprovide a range of gray levels large enough to probe thedifferent transitions between gray values. At the sametime, it needs to provide a statistically known representa-tion that allows the generation of a large number of ran-dom realizations required to conduct a computationalreader study with high statistical power (over 200 inde-pendent realizations per experimental condition for thework described in this article). An additional desiredproperty of the choice of background is that it mimics insome fashion the anatomical structures found in clinicalimages from volumetric systems. A mathematical back-ground formulation that provides all of these criteria is pres-ently not available. In spite of this, we found that choosing astatistically defined background that has been extensivelystudied in two dimensions can, in this case, be extended to3D without introducing a radically different formulation ofthe image properties. The 3D extension to the clustered-lumpy background (CLB) has the additional benefit that ithas been shown to mimic mammographic anatomical struc-tures in its two-dimensional implementation. In fact, it canbe shown that the projection through a 3D CLB yields atwo-dimensional CLB with related characteristics in terms ofthe parameters of cluster and lump size and density. It canbe further argued that if the individual projections of a breasttomosynthesis device have similar properties as regular
mammograms and therefore can be simulated with a two-


dimensional CLB, the object that was used to generate theprojections can be described in a simplified manner, by a 3DCLB for which the parameters need still to be defined.

We recognize that employing a 3D CLB will not beuseful in obtaining modality-specific results concerningslow displays. In spite of this limitation, the 3D CLBscan be used to investigate the more fundamental as-pects of slow displays in a more controlled manner.Although more realistic anatomic images might providedata that is more applicable to a clinical environment,it will also complicate the visual task with respect toanatomical landmark cueing and additional informationbeyond the detection task.

Because of these reasons, we use synthesized 3DCLB with signals added to selected slices as our rawimages for the study. Synthetic cluster lumpy back-grounds are believed to be a good facsimile of realistictwo-dimensional mammographic backgrounds and sta-tistically equivalent up to the second order (15). TheCLB is generated by first producing a random uniformdistribution of delta functions called cluster centers(K). Then each center is used as the spatial origin for asecond process of placing a random distribution (Nk) ofdelta functions to replace those in the first step. K andNk are sampled for Poisson distributions. The deltafunctions are later replaced with analytically definedblobs. The clusters and blobs are both distributed in3D, resulting in a CLB cube. The number and distribu-tion of clusters and blobs and the size of the blobs canbe chosen to generate backgrounds with different sta-tistical properties (see Fig 2 for examples). When thereare fewer clusters and blobs, or blob sizes are small,the background images have sparse structures and thereare more high spatial frequency components in the im-age. As the number of blobs increase, the images lookmore like filtered white noise, and as the size of theblobs increase, the images become flatter.

The 3D CLB described constitutes a powerful tool forthe assessment of a variety of 3D imaging modalities us-ing a statistically defined background with 3D correla-tions. Correlated slices can be selected from the cube anddisplayed in stack-mode on medical displays or fed tocomputational observers. We implemented this approachand generated cubes of 128 � 128 � 128 pixels. Theprocess can be described as follows,

K Nk1

g(r) � �k�1 �n�1 b�akn

(r � rk � rkn), R�kn� (1)

where rk represents position of the kth cluster, akn rep-resents scale of the nth blob within the kth cluster, �kn

is rotation angle of the nth blob within the kth cluster,and b(r,R�) is blob profile rotated at angle �. To repre-sent the power-law characteristics of the background,asymmetrical exponential blobs are used with charac-teristic length Lx and Ly in x and y directions. The de-tails can be found elsewhere (15). In this work, blobs andclusters are randomly distributed in a 3D cube to simulate3D breast CT or tomosynthesis slices. Fifty independent 3DCLBs are generated for the temporal effect calculation. Themean number of clusters was set to be 160, and the meannumber of blobs per cluster was set to be 20. The designernodule is used as a signal to be detected (16) and added tothe center slice.

Slices are run through the temporal response model ofmedical LCDs. To represent the changes in the luminancefields as seen on the display screen, we developed amethod based on the measured gray level transitionpercentage matrix P(g1, g2). In this method, the imagesequence is computed using the percent of the targetluminance achieved after a frame time. The luminanceof pixel (i, j) in an image sequence seen in the screen at aparticular time index n (denoted by L(n)) is given by:

L(n) � L(n � 1) � �L(g0(n), C�1(L(n � 1))) * P(g0(n),

C�1(L(n � 1))) (2)

where g0(n) is the target gray level of pixel (i, j) at timeindex n, L(n � 1) is the achieved luminance level at theend of frame n�1. The achieved gray level correspondingto an arbitrary screen luminance is calculated using theluminance response function of the device (L � C(g)).Because actual luminance values are mapped to a float-ing-point representation of the gray scale, we defineC�1(L) as its nearest integer value. �L(*) is the theoreti-cal increase in luminance given by the grayscale calibra-tion of the display device for the corresponding gray lev-els, and denotes the element-wise multiplication of matrixelements. For the next time index or frame, a similarequation can be written:

L(n � 1) � L(n) � �L(g0(n � 1), C�1(L(n))) * P(g0(n � 1),

C�1(L(n))) (3)

The change of luminance at pixel (i,j) is illustrated in
Fig 3. Note that the luminance level depends not only
373


Figure 2. The three-dimensional CLBs can be generated with different spatial frequencies and contrasts by using different numbersand distributions of clusters, blobs, and different blob sizes. K is the number of clusters, Nk is the number of blobs, and Lx, Ly are thedimensions of the blobs. The images are normalized to a maximum gray level of 225. (a) K � 50, Nk � 10, Lx � 2, Ly � 1. (b) K � 50,Nk � 10, Lx � 3, Ly � 2. (c) K � 50, Nk � 10, Lx � 4, Ly � 2. (d) K � 100, Nk � 10, Lx � 2, Ly � 1. (e) K � 100, Nk � 10, Lx � 3, Ly �2. (f) K � 100, N � 10, L � 4, L � 2 (g) K � 200, N � 10, L � 2, L � 1 (h) K � 200, N � 10, L � 3, L � 2. (i) K � 200, N � 10,
k x y k x y k x y k
Lx � 4, Ly � 2.

374


on the temporal characteristics of the display device,but also on the history of the pixel luminance. The cal-culations assume that luminance is constant in a frametime, shown as the horizontal thick lines in the graph.In reality it will take some time for the luminance tochange from frame n �1 to n. The luminance may takedifferent tracks depending on different transitions anddisplay technology, as illustrated in Fig 1 for transitionat frame n. We also consider the linear case to com-pare with the constant luminance case. The linear ap-proximation assumes that the integrated luminance is agood estimate for the perceived luminance. Alterna-tively, a constant luminance model can also be used.By running the CLB slices through such an LCD tem-poral response model, images that represent actual dis-played values on LCDs can be calculated. Figure 4shows an example of our method to simulate temporalresponse effects on medical images.

Luminance contrast between signal and background isone figure of merit used to judge image quality, which is

Figure 3. Pixel luminance change at different frames. L(n � 1) isthe achieved luminance of pixel (i, j) at frame n � 1, correspond-ing to a gray level g(n � 1), which can be obtained from the dis-play luminance curve (L(n) � C(g(n))). At frame n, the target graylevel is g0(n) (target luminance level L0(n)). Because of slow tem-poral response, the achieved luminance at frame n is L(n). Insome cases, the luminance at the end of a frame is taken as theperceived luminance. If the luminance transition within a frame isconsidered, the perceived luminance can be approximated to bethe average luminance in the frame as shown by �L.

defined as:

LC �Ls � Lb

Lb

(4)

where Ls is the average luminance of the signal and Lb is theaverage luminance of the background in the region adjacentto the signal. However, this metric doesn’t always correlatewith detectability. To further quantify the effects of temporalresponse, a computational observer was used to estimatedetectability under different browsing speeds.

In this work, we used a CHO that incorporates the con-trast sensitivity of the human visual system. We call thisobserver a contrast-sensitive CHO (csCHO). In previouswork, we have compared the performance of the csCHOwith the average performance of 10 human observers from apsychophysical study by Park et al (10). A choice of chan-nels for the csCHO is difference-of-Gaussian channels (seeAppendix for details).

RESULTS

Two medical LCDs were measured for this study: a 5million pixel monochromatic display with an array of2,048 � 2,560 pixels and a 1 million pixel color display.We chose these display systems to represent a high-perfor-mance LCD monitor that could be used in high-resolutiondisplay for a breast tomosynthesis system and a consumer-grade LCD. As we have shown in previous work (20), thetemporal characteristics of high-performance 5 million pixelmonochromatic display systems is similar to 3 million pixelsystems used as PACS monitors. Measured response timesranged from 15 milliseconds to 150 milliseconds (Fig 5) andare dependent on temperature. One hour of warm-up reducesthe response time by a factor of 2. We found that the re-sponse is better for large gray scale transitions and is worsewhen the starting gray level is between 10 and 20 for a tar-get between 25 and 150.

Figures 6 and 7 illustrate the achieved luminance afterone to four frames for a 1 million pixel color display and a5 million pixel monochromatic display LCD. Each point inthe figures represents the percentage of luminance achievedafter n frames for different gray level transitions when the x,y axes correspond to the starting 217 (from) and ending (to)gray levels. Ideally, 100% of luminance will be realized af-ter one frame, but for both displays, a significant number ofthe transitions remain below 80% after four frames. When a
stack of images is displayed on these LCDs, different brows-
375

d a slow (right) response device are shown.


ing speeds will lead to different degradations of image quali-ties. For transitions in which the temporal response is poor,the luminance after one and four frames reaches only 12%and 75%, respectively.

Figure 8 shows the luminance contrast and detect-ability reduction as the browsing speed is increased.The results are for CLBs generated with an averagegray level of 32 and 64. The signal has a peak graylevel of 32 and is added to the background of the cen-ter slice. In a fast display, the images shown in themonitor will be an expected representation of the gen-erated image data. For slow LCDs, the pixel luminancewill be somewhere between the beginning and the tar-get image value. We obtained 50 groups of CLBs usingthe LCD temporal response model (see Eq 2 and 3),and calculated the luminance contrast in the centerslice. The luminance at the end of the frame was takenas the luminance of the entire frame. The squares indi-cate the average results, whereas the black point depictthe variability in contrast reduction over the 50 cases,which is due to the random realization of the back-grounds and the random gray level over which the sig-nal is located. This in turn results in different graylevel transitions that suffer from different transitionspeeds. The plot of contrast versus browsing speedshows that, for a speed of 10 images/second, the reduc-tion can be up to 8%. Luminance contrast decreaseswith browsing speed 50% on average with a maximumreduction in the test set of 70%.

The results with the computational observer (Fig 8)suggest that a reduction in contrast translates into a

Figure 4. Example of a clustered lumpysion of clustered lumpy background (15) athe signal corresponding to a fast (left) an

background used in the study, with 3D exten-nd designer mass (16). The central slice with

reduction in detectability. Decreased signal contrast

376

Figure 5. Temporal response measurements corresponding todifferent devices ((a) 1 million and (b) 5 million pixels) used in thestudy to simulate the effect of the different display temporal char-acteristics on target detection. Contour lines indicate the re-
sponse time for different regions in the plots.

ame


almost always results in decreased detectability. How-ever, increased browsing speed also decreases thenoise. When the signal contrast and the noise bothchange, the detectability might increase or decrease. Inthe case studied, the reduction in signal luminance con-trast is more significant than the reduction of back-ground noise, thus reducing of detectability. Further-more, our study suggests that the magnitude of the ef-fect depends on the gray level of the background andon the gray level difference of the signal.

In the simulation, we assume the pixel luminance ofeach frame is constant and equal to the value achievedat the frame end. For a slow display, the pixel lumi-

Figure 6. Luminance achieved after one, two, three, and four fraget luminance for display 1 million pixel color display: after one fr

nance changes gradually during a frame time (as de-

picted in Fig 1). When the history of the luminance ina frame is considered, the perceived luminance is theaverage luminance during the frame. In such a case,the perceived luminance cannot be directly calculatedusing the percentage matrix, which only reflects theluminance at the end of the frame. Moreover, the lumi-nance change within a frame may take different formsdepending on display technologies and specific imagegray levels considered. To simplify, we assume theluminance changes are linear during the frame time sothat the perceived luminance is the average of the be-ginning and ending luminance. Based on this assump-tion, the calculations were repeated, and the results

expressed as a percentage of the desired transition to the tar-(a), two frames (b), three frames (c), and four frames (d).

mes

shown in Fig 9.

377

e fra


At a browsing speed of 30 frames/second, the aver-age luminance contrast of signal decreased to 30% ofthe original value. The area under the curve for thechannelized observer also decreased under the linearluminance assumption.

DISCUSSION

When the pixel value changes in an image sequence,it is expected that the luminance will change instanta-neously. For a slow LCD monitor, it usually takes sev-eral frame times for the pixel luminance to reach the

Figure 7. Luminance achieved after one, two, three, and four framluminance for display 5 million pixel monochromatic display: after on

target luminance, resulting in degradation of the image

378

quality of fast changing image sequences. We foundthat the speed for different gray level transitions is dif-ferent. For some of the transitions, even four framesare not enough for a full luminance change, as shownin Figs 6 and 7.

When a stack of images is displayed on an LCD, theluminance may or may not reach the target value dependingon the browsing speed. Luminance change within a framedepends on the LCD technology and also on the specificgray level transitions considered. When assuming the lumi-nance is constant in a frame time and equal to the frame-endluminance, we found less degradation than when assuming a

pressed as a percentage of the desired transition to the targetme (a), two frames (b), three frames (c), and four frames (d).

es ex

linearly changing luminance. The linear luminance model

sede mo


attempts to account for the continuous change in luminanceexpected within a frame in a simple way that does not spe-cifically account for the electrical, optical, and mechanicalphysics that govern the LCD display. Consequences of thesimplification may overestimate the image quality degrada-tion or may be too optimistic depending on different shapes

Figure 8. Reduction in contrast (a) and in detectability (b), expres1 million pixel and 5 million pixel displays with a constant luminanc

of luminance transitions curves.

The result from the computational observer studyindicates that the detectability defined by area underthe curve tracks the trend observed for luminance sig-nal contrast. Our computational observer model takescontrast sensitivity of human vision into account butdoes not yet include the temporal effects in the contrast

as the area under the receiver operating characteristic curve, fordel.

sensitivity. As a result, in this study we did not include

379

sedtrans


the human vision system’s temporal response, whichmight also have an effect on the perceived images athigh browsing speeds.

One fact to point out is that this study aims at un-derstanding temporal effect of displays on medical im-age perception. We choose to use csCHO, which pre-

Figure 9. Reduction in contrast (a) and in detectability (b), expres1 million pixel and 5 million pixel displays, with a linear luminance

dicts human performance well. However, for the signal

380

type in this article, the performance of CHO andcsCHO, when normalized, equally predict the task per-formance.

One of the main limitations of this approach withrespect to the choice of model observer is that we areaddressing a 3D effect with a two-dimensional ob-

as the area under the receiver operating characteristic curve, forition model.

server. In other words, we are limiting the observer to


investigating a single slice in the stack. An alternativeapproach is to use an observer that combines the re-sults from all the slices into a single decision about thepresence or absence of the signal. This approach, simi-lar to and inspired by work done by Eckstein et al(18), might provide additional insight into the temporaleffects of image browsing. However, this approach alsofails to incorporate another component missing in ouranalysis; the temporal response of the human visualsystem. Literature suggests that the visual system actsas an integrator with a time constant in the order ofseveral tens of milliseconds (19). Given a long integra-tion time, our assumptions of using an average lumi-nance within a frame might not be completely unrea-sonable. Still, a significant limitation remains in that itis highly unlikely that the display refresh cycle and thehuman eye integration are synchronized.

Another issue related to the human vision systemthat might play a significant role in dynamic visualtasks is the ability to detect motion, defined in thiscase as a sudden appearance of a large-contrast featurein the image. Radiologists rely heavily on detecting the“pop-out” effect of abnormalities as frames arebrowsed by. The degradation of luminance images be-cause of slow display is precisely what affects the per-ception of this pop-out effect and thus warrants furtherconsideration. One way to address this issue with re-gards to human reader experiments is to cue the ob-server so that the reader knows where the signal mayappear in spatial and temporal coordinates. In this way,the observers will not be looking for changes at everypoint in the temporal sequence but will focus on a par-ticular range of time with the help of a reference signal(maybe displayed in an additional box situated next tothe signal-absent or signal-present image sequences).This is analogous to the disclosure of the signal sizeand profile to observers while doing a two-alternativeforced-choice experiment. These issues are being inves-tigated in our current work which we expect to publishsoon.

Another limitation of our approach involves the choice ofbackground. The main limitation of using the 3D CLB isthat it does not represent any specific background imagefound in clinical systems. However, its statistical definitionallows generating a large number of independent randomrealizations which are needed to perform the reader studywith large statistical power. The 3D CLB strategy pro-vides means of adapting the statistics of the back-
ground to reflect data that has different content of
low-, mid-, and high-spatial frequencies. This will be-come useful when investigating the additional compli-cation of varying signal size on the interrelationship oftemporal and spatial effects. This topic will be ad-dressed in a subsequent publication. Another reason forworking with a 3D CLB is related to the fact that thegray level range in the image can be adapted to repre-sent a case in which some gray level transitions aremore frequent. This is of particular importance whenconsidering that the temporal characteristics of displaysbased on liquid crystal technology are not uniformacross gray levels, but instead, have a strong depen-dency (up to two orders of magnitude) between differ-ent transitions (20). The adaptation of a clinical back-ground for a controlled study like the one described inthis article seems not as straight forward.

A final consideration for the study is that the signal used,which is present in one slice, could be thin relative to slicethickness. To better simulate medical images, a signal thatappear in multiple slices should be considered. A multiple-frame signal would allow better detection of the signal by theobserver, with several slices contributing to a higher detectabil-ity. This methodology will be addressed in a follow-up article.

CONCLUSIONS

The effects of slow LCD response were explored usingsimulated images based on measurements of display per-formance. Using our simulation method and a contrastsensitive CHO, we investigated how temporal response ofthe LCD device affects observer performance in detectiontasks using complex backgrounds. We found that the slowtemporal response of displays degrades the performanceof observers in stack-mode readings of volumetric imagesby reducing the effective luminance contrast of lesions.Our results for the model observer will be validated withhuman studies in a follow-up article. Slow temporal re-sponse of the display device could lead to limits for therendering speed of large volumetric image datasets (fromCT, magnetic resonance, or breast tomosynthesis). Ourmethod can be used to quantify the effect of differentdisplay temporal characteristics on a variety of visualtasks. However, more model observer calculations andvalidation with human readers are needed to developguidelines for acceptable image browsing speeds. Inparticular, validation with more realistic backgrounds
needs to be carried out before establishing conclusive
381


recommendations for modality-specific imaging inter-pretation.

REFERENCES

1. A. Badano. AAPM/RSNA tutorial on equipment selection: PACS equip-ment overview: display systems. Radiographics 2004; 24:879–889.

2. Tsukuda T. TFT/LCD: liquid crystal displays addressed by thin filmtransistors. Amsterdam, The Netherlands: Gordon and Breach SciencePublishers, 1996.

3. Kurita T, Saito A, Yuyama I. Consideration on perceived MTF of hold typedisplay for moving images. Proc Int Display Workshop’98 1998; 823–826.

4. Miseli J. Motion artifacts. Proc SID 2004; 86–89.5. Mathie AG, Strickland NH. Interpretation of CT scans with PACS image

display in stack mode. Radiology 1997; 203:207–209.6. Reiner BI, Siegel EL, Hooper FJ, et al. Radiologists productivity in the

interpretation of CT scans: a comparison of PACS with conventionalfilm. AJR Am J Roentgenol 2001; 176:861–864.

7. den Boer W. Active Matrix liquid crystal displays: fundamentals andapplications. Burlington, MA: Elsevier, 2005.

8. Nakamura H, Crain J, Sekiya K. Optimized active-matrix drives for liq-uid crystal displays. J Appl Phys 2001; 90:2122–2127.

9. Barrett HH, Myers KJ. Foundations of image science. Hoboken, NJ:Wiley, 2004.

10. Park S, Gallas BD, Badano A, et al. Efficiency of the human observerfor detecting a gaussian signal at a known location in non-gaussiandistributed lumpy backgrounds. J Optical Soc Am A Optics Image SciVision 2007; 24:911–921.

11. Myers KJ, Barrett HH. Addition of a channel mechanism to the ideal-observer model. J Opt Soc Am A 1987; 4:2447–2457.

12. Abbey CK, Barrett HH. Human- and model-observer performance inramp-spectrum noise: effects of regularization and object variability.J Opt Soc Am A 2001; 18:473–487.

13. Gallas DB, Barrett HH. Validating the use of channels to estimate theideal linear observer. J Opt Soc Am A 2003; 20:1725–1738.

14. Liang H, Badano A. Precision of gray level response time measure-ments of medical liquid crystal display. Rev Scient Instr 2006;77:065104.

15. Bochud FO, Abbey CK, Eckstein MP. Statistical texture synthesis ofmammographic images with clustered lumpy backgrounds. Optic Exp1999; 4:33–43.

16. Burgess EA, Jacobson FL, Judy PF. Human observer detection experi-ments with mammograms and power-law noise. Med Phys 2001;28:419–437.

17. Barten PGJ. Contrast sensitivity of the human eye and its effects onimage quality. Bellingham, WA: SPIE Press, 1999.

18. Eckstein MP, Whiting JS, Thomas JP. Role of knowledge in human vi-sual temporal integration in spatiotemporal noise. J Opt Soc Am A1996; 13:1960–1968.

19. Bartlett NR. Thresholds as dependent on some energy relations andcharacteristics of the subject. In Graham CH, ed. Vision and visual per-ception, New York: Wiley, 1965; 154–184.

20. Liang H, Badano A. Temporal response of medical liquid crystal dis-plays. Med Phys 2007; 34:639–646.

21. Park S, Badano A, Gallas BD, et al. A contrast-sensitive channelized
observer. IEEE Trans Med Imaging. In press.
382

APPENDIX: COMPUTATIONALMODEL OBSERVER

The luminance response curve of the display from thepsychophysical study is used to convert image values toluminance values. The Barten model (17) is employed toincorporate the contrast sensitivity of the human visualsystem into the CHO with difference-of-Gaussian chan-nels. The main mechanism underlying this observermodel determines whether a pixel gray level differencefrom the region of interest mean is larger than the con-trast threshold. First, we calculate the contrast sensitivitythreshold (C) as follows:

C(l�) �1

C0csf(�, l�)(5)

where � is the frequency of the signal to be detected, �l isthe mean luminance in a field of view (FOV) of an im-age, and csf(.) is the contrast sensitivity model of the hu-man visual system. Then we use C( �l) to determine if apixel luminance is above the threshold:

lk � l�

l�� C(l�) → lk � l� (6)

where lk is the luminance of the kth pixel in the image.If the given relation is satisfied, the luminance of thekth pixel in the image is replaced by the mean lumi-nance of the field of view; if not, the luminance of thekth pixel is left unchanged. Finally, the CHO with dif-ference-of-Gaussian channels was applied to the result-ing image. We fixed C at a value for which the csCHOmatches human performance in one of the experimentconditions. Then we calculated the csCHO performancefor the other experiments. csCHO has been shown towell predict human performance in detection tasks us-ing lumpy backgrounds. More details on the perfor-mance of the csCHO will be presented in a separate
article (21).

Documents

Image Browsing in Slow Medical Liquid Crystal Displays