Application of Pattern Recognition Techniques for the Analysis of Histopathological Images-libre

Application of Pattern Recognition Techniquesfor the Analysis of Histopathological Images

Adam Krzyzak, Thomas Fevens, Mehdi Habibzadeh, and ukasz Jelen

Abstract. In this paper we discuss applications of pattern recognition and imageprocessing to automatic processing and analysis of histopathological images. Wefocus on two applications: counting of red and white blood cells using microscopicimages of blood smear samples and breast cancer malignancy grading from slidesof fine needle aspiration biopsies. We provide literature survey and point out newchallenges.

Keywords: CBC, microscopic medical images denoising, binarization, segmenta-tion, edge preservation, granulometry, fine needle aspirates, breast cancer malig-nancy grading.

1 Introduction

Automatic detection of pathologies from histopathological images is currently veryactive and important area of research. In the present paper we will survey two ap-plications of pattern recognition and image processing in this emerging field: au-tomatic processing of blood smear images and automatic grading of breast cancerfine needle biopsy slides. The paper is organized as follows: in sections 2-5 wewill review processing of blood smears and in sections 6-9 we will focus on cancergrading.

Adam Krzyzak Thomas Fevens Mehdi HabibzadehDepartment of Computer Science and Software Engineering, Concordia University,1455 De Maisonneuve Blvd. West, Montral, Qubec, Canada H3G 1M8e-mail: {krzyzak,fevens,me_hab}@encs.concordia.caukasz JelenFaculty of Life Science and Technology, Wrocaw University of Environmental andLife Science, Norwida 2527, 50-375 Wrocaw, Polande-mail: [email protected]

R. Burduk et al. (Eds.): Computer Recognition Systems 4, AISC 95, pp. 623644.springerlink.com Springer-Verlag Berlin Heidelberg 2011

624 A. Krzyzak et al.

2 Manual Analysis of Blood Smear Images

Analysis of microscopic medical images is an important interdisciplinary probleminvolving both physicians and computer scientists. One of the important and activeareas of research is the problem of counting blood cells (CBC) [1, 2, 3] which isused as screening test to check such disorders as infections, allergies, problems withclotting, and it helps diagnosing and managing a large number of diseases. In prac-tice a panel of tests is carried out that examine different blood components such ascounting white blood cells (WBC) [2, 4], white blood cells differential, counting redblood cells (RBC) [2], checking for signs of disease and the counting the number ofinfected cells. Blood cell counting and blood film examination are widespread diag-nostic techniques [3]. A blood smear is obtained by drawing blood from a vein andplacing a drop on a glass slide [3]. The blood film is stained [3] using e. g., Wrights,Giemsa, or May-Grunwald staining techniques and imaged with a transmission lightmicroscope. The definitive diagnosis of blood smear infection is done by manuallyfinding disorders and abnormalities in blood films through a microscope, countingblood smear particles and cells with disorders, which are not only a time consumingtask but also prone to human error. The erythrocytes and leukocyte types that thecurrent equipment is able to manage are restricted to few classes [5] and stainingprocess requires expensive chemicals.

As mentioned, the microscope inspection of blood slides provides importantqualitative and quantitative information concerning the presence of hematic patholo-gies [6], however the number of different sub-cell types that can come out especiallyfor WBC count is relatively large and typically more than 20 [5]. Normal periph-eral blood contains the following types of leukocytes (the numbers in brackets givethe typical proportion of the cell type): segmented neutrophil (40- 75%), lympho-cyte (25-33%), monocyte (2-8%), eosinophil granulocyte (1-4%), band neutrophil(1-3%), plasma cell (0.2-2.8%), basophil granulocyte (0.5%), and atypical lympho-cyte. Other cell types which are observed in certain diseases include: metamyelo-cyte, myelocyte, promyelocyte, myeloblast and erythroblast [3] and this increasesthe difficulty in building a feasible system. This process can be automated by com-puterized techniques which are more reliable and economic. Therefore there is al-ways a need for the development of systems to provide assistance to hematologistsand to relieve the physician of drudgery or repetitive work. So, more systems forautomatic processing of medical images are being developed and during blood filmexamination, the individual types of blood smear particles (leukocytes and erythro-cytes) are enumerated yielding so called differential count.

Our goal is to develop and validate the necessary image and pattern recognitionprocessing algorithms to quantify and detect microscopic particles on slides to en-hance automated system to characterize blood health status of patient. In essencethat will enable us to determine the fast, accurate mechanism of segmentation andgather information about distribution of microscopic particles which are help to di-agnose status of abnormality or normality and represent a factor of combatively andquality for the modern laboratories of clinical analysis.

Analysis of Histopathological Images 625

3 Automatic Processing of Blood Smear Images

During blood film examination, the individual types of blood smear particles (leuko-cytes and erythrocytes) are enumerated and then blood films are usually made toinvestigate hematological problems [1, 2]. The history of research into automatedblood slide examination dates back to 1975, see Bentley & Lewis [7]. However itis only recently that digital photography, computer speed, RAM size and secondarystorage capacity have made automatic blood processing possible. The analysis ofblood slides must be fully automated to be useful [8]. Due to complexity of theproblem at hand (Costin et al. [9]) most of the papers are limited to image-basedcomparisons based on red cells segmented either manually, see Bentley & Lewis[7], Albertini et al. [10], or semi-automatically, see Robin-son et al. [11], Costin etal. [9] and Gering & Atkinson [12].

There is a vast amount of literature dedicated to differential blood counts. A keystep in automating this process is the segmentation of cell boundaries. Initial successon segmentation of medical images was obtained with graph theory (Martelli [13],Fleagle et al. [14], Fleagle et al. [15]) which was used to navigate around edge pix-els found in an image. However this approach has involved images of single objectsmanually located in an image, and does not address the problems of multiple objectsin the image, object location, removal of extraneous edges (internal to the cell), orthe selection of suitable starting and ending points for the graph search. Further-more, thresholding has been used to pre-process images as an aid to segmentation(Gonzalez & Woods [16]). With red blood cell images this causes problems due tothe pale nature of the interior of the cells, which then necessitates further processing.Adjouadi and Fernandez [17] find the cell borders using eight-directional scanningwithin thresholded images of normal blood. The problem with this approach is thatit would not find the whole of the border of severely deformed cells as these containedge points that would not be reached by any of the eight scan-lines. Moreover theprocess does not result in identification of the points within the contours. Di Rubertoet al. [18] follow thresholding with a segmentation using morphological operatorscombined with the watershed algorithm [19]. However their work is aimed at seg-mentation of red blood cells containing parasites and is designed to increase thecompact nature and roundness of the cells. Such assumption of roundness is not ap-propriate for segmentation of all particles for the purpose of classification, becauseblood smears may contain deformed red blood cells or some WBC types [1] whichare not 100% convex and circular. The method is also complicated requiring nineintermediate steps and does not result in border identification. It also requires somepreprocessing which is not always applicable for all possible slides and then thereis no dynamic way to better control image acquisition.

Another popular automatic approach to border detection is that of active con-tours, or snakes (Kass, Witkin & Terzopoulos [20]), which can be applied either tothe original image or to an edge image. However when used to identify cell borders,the resulting contours do not correspond with the exact borders of the cells (Ongunet al. [21], Wang, He & Wee [22]),which would cause problems with subsequentRBC classification, where the exact boundary shape is important. Other problems


with the use of contours for images of peripheral blood smear slides include theinitial positioning of the multiple contours required; the tendency of the contours tofind the inner pale cell areas in addition or instead of the outer edges; and the fail-ure of contours to identify pixels interior to the contour. Other works using activecontours for tracking boundaries of WBC include Ongun [23] but it could not dealwith WBC overlapping problem. Lezoray [24] proposed a region-based WBC seg-mentation strategy using seed flooding. However, it relied greatly on the proper seedextraction using prior knowledge of color information. Kumar [25] defined a newedge operator and tried to get precise nucleus edge. But it required relatively weakedge existing between red blood cell (RBC) and background, which was often miss-ing. An automated system where cells are segmented using active contour models(snakes and balloons, initialized by morphological operators) are presented in [26].The shape and texture features are used for classification. A two step segmentationprocess is used by Sinha and Ramakrishnan [27]. First the HSV transformed imageis clustered using k-means followed by an EM-algorithm. The shape, color, and tex-ture features are then used in a neural network classifier. A mean-shift-based colorsegmentation procedure applied to leukocyte images is described in [28]. Segmen-tation is performed in the L*u*v* color space. A watershed-based segmentation isused in [29]. First a sub-image containing a leukocyte is separated from the cell im-age. The nucleus region is then detected by scale-space filtering and the cytoplasmregion by watershed clustering of the 3-D data. WBC classification in recent workHamghalam et al. [30] utilizes Otsus thresholding method to segment nuclei. Theresults are independent of the intensity differences in Giemsa-stained images of pe-ripheral blood smear and active contours are used to extract precise boundary ofcytoplasm.

As mentioned previously, the nature of microscopic particles is not simple andautomatic processing of images in medicine is a complicated task. This is becausesome of the basic tasks to be performed such as pre-processing, segmentation, clas-sification, object recognition and inference require extensive understanding of thespecific problem. This requires comprehensive knowledge in many disciplines suchas medicine, computer science, image and signal processing.

4 Methodology and Algorithms

In the next few sections we will review basic steps for processing microscopic bloodsmear images.

4.1 Image Acquisition and Denoising

The first step is to convert RGB channels images to the green channel as it is morereliable than the red or blue channels for noisy and distorted images. The next stepis choosing an effective denoising tool. To design a reliable automated segmenta-tion system that may be used under different conditions such as a variety of mi-croscopic staining techniques, types of chemical materials used, microscope types,


illumination conditions, human error, etc., a pre-processing step is required. Theaccuracy of this stage affects the system performance. There are wide variety tech-niques for enhancing image quality. One of the most practical and widely used de-noising technique is wavelet shrinkage approach which thresholds the wavelet coef-ficients of an image. Wavelet coefficients having small absolute value are consideredto encode mostly noise and very fine details of the signal. In contrast, the importantinformation is encoded by the coefficients having large absolute values. Removingthe small coefficients and then reconstructing the signal could produce signal withlesser amount of noise. The biggest challenge in the wavelet shrinkage approach isfinding an appropriate threshold value [31].

The wavelet shrinkage approach can be summarized as follows:

1. Apply the wavelet transform to the signal.2. Estimate a threshold value.3. Remove (zero out) the coefficients that are smaller than the threshold.4. Reconstruct the signal (apply the inverse wavelet transform)In [32, 33] Daubechies wavelet with soft thresholding and Bivariate Shrink togetherwith PSNR ratio has been used. In using soft thresholding based on following con-cepts the user should calibrate the parameters of the algorithm. The optimal thresh-olding obtained by using soft thresholding which depends on experience and on thetype of images.

In Figs. 1 and 2 we illustrate wavelet denoising including two simulation studies:one for images corrupted by moderate additive normal noise with deviation 30 andthe second for highly corrupted by additive normal noise with deviation 100.

Fig. 1 a) original image (red channel); b) noisy image; c) median denoising; d) soft thresh-olding denoising; e) Bivariate denoising.


Fig. 2 a) noisy image; b) median denoising; c) soft thresholding denoising; d) Bivariate de-noising.

Table 1 PSNR levels for various denoising techniques for images with moderate and highnoise.

Additive Noise deviation30 100

Noisy Image 19.2149 10.4516PSNR Denoised Image using Median 25.5666 16.5183

Denoised Image using Thresholding 23.5460 18.3421Denoised Image using Bivariate 27.6236 20.4822

For moderate noise and high noise, the PSNR experimental results are summa-rized in Table 1. From the experimental results it can be concluded that for moder-ate noise the Bivariate Shrink filter produces best results. It produces the maximumPSNR for the output image compared to the other filters. However, Bivariate out-put, is somehow blurred and some post-processing involving de-blurring and edgepreserving may be needed. For images heavily corrupted by noise with low PSNRvalue (10.4516) the Bivariate Shrink filter is again best. It produces the maximumand acceptable PSNR for the output image compared to the other filters. It can alsobe observed that for high noise levels soft thresholding produces better results thanthe classical median filter.

4.2 Edge Preservation

The aim of the next edge preservation step is to recover degraded and blurred imageswhile reducing the negative effects of noise such as blurred edges produced by theBivariate Shrink filter. This step can serve as preliminary step prior to binarization


and object segmentation. Different edge preservation techniques have been used inpractice. They include median filter [34], symmetrical nearest neighbor (SNN) filter[35], convolution kernel filters [36], preserving color reduction method [37], bilat-eral techniques [38], and the Kuwahara filter [39]. In computer simulations we havelearned that Kuwahara filter works best. This can be justified by intrinsic character-istics of microscopic particles for which Kuwahara filter yields the sharpest edgeswhich leads to better binarization in next step (see Fig. 3). However, the output maybe somewhat toothy and jagged.

a) b) c) e)d)

Fig. 3 a) edge preservation with Bilateral, b) convolution kernel, c) EDGEPS [37], d) SNN,e) Kuwahara filters, respectively.

4.3 Binarization

After denoising and edge enhancement, binarization is the third step which allowsto extract some features, having sub images and get ready to apply new techniquefor different purpose over the images. Generally, binarization methods can be ap-plied with global and local thresholding. Different binarization methods includethe approaches of Niblack [40], Bernsen [40], Sauvola [41] and Otsu [42]. Com-puter experiments with different samples and initial conditions (see Fig. 4) showthat Niblack approach is the most reliable method to maintain disjoint componentswhich is crucial in avoiding over or under segmentation.

Mehdi et al [43] proposed a modified binarization method that merges Niblackand Otsu approaches. This process reduces limitations and drawbacks of each ofthem. Niblack uses local thresholding based on average and standard deviation ofa local area. The size of the window must be large enough to suppress the noisebut at the same time small enough to preserve local details. In practice, a window

a) b) c) d)

Fig. 4 Binarization methods: a) Bernsen; b) Sauvola; c) Otsu; and d) Niblack


size of 1515 works well in all available image databases. The Niblack methodtends to result in overlapping objects that are too close to one another which in turnleads to false segmentation results. In the modified version, pixels are labeled asbackgrounds pixels if they are labeled as either background pixels in Niblack or inOtsu and the remaining pixels are kept as foreground pixels (objects). Using thismerging process, we mitigate the problem of extra small spurious regions producedby the Niblack algorithm.

4.4 Size Estimation

Binarization and some post-processing to enhance the quality of binary image is fol-lowed by feature extraction which helps to differentiate various types of particles inthe image. A normal blood cell is one of two major particles: a RBC with a normalprobability distribution function (PDF) with average size around 6.0-8.5 m or aWBC with average size around 7-18 m which includes a nucleus and cytoplasm isabout 1-3 times bigger than normal and mature RBCs. We use size characteristics asan effective factor to distinguish between the two main types of cells. Granulometry[44] can determine the size distribution of image objects without explicitly segment-ing each object first. According to normal blood PDF and RBC to WBC ratio, themaximum regional peak in pattern spectrum diagram correlates to the number ofRBCs with an acceptable RBC radius size. We summarized the granulometry algo-rithm in the next section.

Granulometry Algorithm

Granulometry is concerned with size distribution of cells in binary images. It usesstructure elements which are morphologically dilated to the maximum size and ap-plied to the image. The shape of structure element depends on the type of objectsunder processing. During the process granulometric density function is determined[7, 8].

Granulometric algorithm starts by applying opening morphology along with de-fined structure element (SE). In normal blood smear images, all available particlesare approximately circular. Hence, we select (disk) shape as default and basic struc-ture element for granulometric algorithm. In ideal output, we expect only one peakfor a single complete circle, but the incomplete circular object shown in Fig. 5 pro-duce local maxima. We call this undesirable effect an edge fracture. We just observethat after applying the edge detection and skeletonisation algorithms to real cellimages which are typically not complete curves the observed circular pieces areregarded as a new objects surrounded between two ideal complete circles. Conse-quently we can expect in granulometric output at least two local regional peaks. Bythis simple work, we find that blood smear particles are not complete circular objectand there are always discrete components on curve tracer, which is another reasonfor undesirable local maxima.

Overall, applying granulometry to RBCs images in normal blood smear can bevery reliable in determination and estimating their size. But for abnormal samples


Fig. 5 Granulometry over simple circle

with different shapes or with extra overlapping between the particles granulometricapproach may fail.

4.5 Segmentation

There are different methods which are directly or indirectly subjected to separationand segmentation objects in disjoint images such as active contours and watershed.Typically watershed is incorporated into the immersion and toboggan methods [45].The accuracy and efficiency of watershed segmentation over images is directly re-lated to the previous steps. In practice, watershed algorithm works best for smoothconvex objects that dont overlap too much. It cannot be an efficient approach in allmicroscopic images with extra overlapping which can happen for some diseases.

5 Conclusion: Segmentation of Microscopic Imagery

In this paper, a simple and step-by-step efficient algorithm has been presented to-gether fully automated detection and segmentation of microscopic imagery. Exper-imental results indicate that the current analysis is accurate and offers remarkablesegmentation accuracy. The performance of the proposed method has been evaluatedby comparing the automatically extracted particles with manual segmentations andother traditional techniques [43]. Furthermore, the introduced method being simpleand easy to implement is best suited for biomedical applications in clinical settings.

6 Cancer Cells Grading

Automatic cancer grading is a very challenging task due to large variation in cancerimaging and analysis. In the remainder of the paper we shall focus on automaticmalignancy grading of breast cancer fine needle aspiration biopsies.

7 Breast Cancer Diagnosis

According to statistics breast cancer is one of the most deadly cancers amongmiddle-aged women. Based on the data provided by the Breast Cancer Societyof Canada about 415 women will be diagnosed with breast cancer each week inCanada. Most of the diagnosed cases can be fully recovered when diagnosed at an


early stage. Cancers in their early stages are vulnerable to treatment while cancersin their most advanced stages are usually almost impossible to treat. During thediagnosis process, the cancer is assigned a grade that is used to determine the ap-propriate treatment. Successful treatment is a key to reduce the high death rate. Themost common diagnostic tools are a mammography and a fine needle aspirationbiopsy (FNA). Mammography, which is a non-invasive method, is most often usedfor screening purposes rather than for precise diagnosis. It allows a physician tofind possible locations of microcalcifications and other indicators in the breast tis-sue. When a suspicious region is found, the patient is sent to a pathologist for a moreprecise diagnosis. This is when the FNA is taken. A fine needle aspiration biopsy isan invasive method to extract a small sample of the questionable breast tissue thatallows the pathologist to describe the type of the cancer in detail. Using this methodpathologists can very adequately describe not only the type of the cancer but also itsgenealogy and malignancy. The determination of the malignancy is essential whenpredicting the progression of cancer.

In this section we will review the computerized breast cancer diagnosis, which isa very active field of research (see sec. 8). Additionally, we will also look over theless active field which is a computerized breast cancer grading (see sec. 9).

8 ComputerAided Breast Cancer Diagnosis

Breast cancer diagnosis is a very wide field of research studying not only medicalissues but also computer science issues. Breast cancer diagnosis is a multi-stageprocess that involves different diagnostic examinations.

Pattern classification is a wellknown problem in the field of Artificial Intelli-gence concerned with the discrimination between classes of different objects [46].We can use the same techniques in cancer diagnosis to assist doctors with their de-cisions. Cheng et al. [47] provided an extensive survey on automated approaches inmammograms classification and importance of computer assisted diagnosis. Sincemammography is one of the preliminary tests performed to locate abnormalities inthe breast tissue, it is used for screening purposes and has raised a lot of interestwithin the scientific community [47, 48, 49, 50, 51, 52, 53].

To the best of our knowledge, the computerized breast cytology classificationproblem was first investigated by Wolberg et al. in 1990 [54]. The authors describedan application of a multi-surface pattern separation method to cancer diagnosis. Theproposed algorithm was able to distinguish between a 169 malignant and 201 be-nign cases with 6.5% and 4.1% error rates, respectively depending on the size ofthe training set. When 50% of samples were used for training, the method returneda larger error. Using 67% of sample images reduced the error to 4.1%. The sameauthors introduced a widely used data-base of pre-extracted features of breast can-cer nuclei obtained from fine needle aspiration biopsy images [55]. Later, in 1993,Street et al. [56] used an active contour algorithm, called snake for precise nucleishape representation. The authors also described 10 features of a nucleus used for


classification. They achieved a 97.3% classification rate using multi-surface methodfor classification.

The features described by the authors are mainly geometrical features of the nu-cleus. These features are:

Radius defined as an average of the radial line segments lengths from the cen-troid of the nuclei to the snake points on the boundary.

Perimeter is the length of the boundary of a polygon connecting snake points. Area is a number of pixels inside the closed snake curve. Compactness = perimeter

2

area

Smoothness of a nuclei contour defined as an average difference between thelength of a radial line and the mean length of the lines surrounding it.

Concavity a measure of nucleus concavity. This is performed by drawingchords between non-adjacent snake points and measuring the extent to whichthe boundary of the nucleus lies on the inside of each chord. The length of thechord that is outside of the nuclei is considered as a measure of the concavity.The concavity is larger when the length of the exterior chord increases.

Concave points measures number of concavities and not their magnitude. Symmetry Here, the major axis (longest chord through the center) is found.

Next, length difference between lines perpendicular to major axis to nucleusboundary in both directions are measured.

Fractal Dimension of a cell this is approximated using a coastline approxima-tion method. Authors measure the perimeter of the nucleus using increasinglylarger segments. Next, they plot the obtained values on a log scale and calcu-late the downward slope which gives an approximation to the fractal dimension.Higher values of the feature provide higher probability of malignancy.

Texture authors define texture as an average gray scale intensity of the nucleus.

Based on the above features, Street [57], in his PhD Thesis, introduced a systemcalled XCyt, that was later improved and described in 2000 [58]. In 1999, Lee andStreet [59] described an iterative approach for automated nuclei segmentation as anaddition to the previously described framework. In 2003, they introduced flexibletemplates to their iterative Generalized Hough Transform approach for segmenta-tion. They created a set of predefined templates of a nuclei and each iteration shuf-fles the templates in such a way that those that were used the most often during theprevious iteration are visited first to save time. The authors were able to segmentnuclei with 78.19% accuracy [60]. They also introduced a neural network approachfor classification stage, achieving 96% accuracy. Classification was based on thefeatures previously described by Street et al. [56].

All work presented above was based on the Wisconsin Breast Cancer Database(WBCD) introduced by Mangasarian et al. [55]. This data-base consists of pre-extracted nuclear features and is widely used among researchers. Features includedin the data-base are the features proposed by Street et al. [56]. WBCD [55] and itsvariations [61, 62] are the only data sets publicly available. Therefore, the majorityof work in this field is performed on this data-base and involves research on differentclassification algorithms.


In 1998, Walker et al. [63, 64] introduced Evolved Neural Networks for breastcancer classification and tested their algorithm on WBCD data-base achieving 96%correctness. Nezafat et al. [65] used WBCD to compare several classification algo-rithms such as k-nearest neighbor classifier, radial-basis function, neural networks,multilayer perceptron and probabilistic neural networks. The authors showed thatamong these classifiers, multilayer perceptron with one hidden layer performed themost efficiently giving 2.1% error rate. Additionally they also compared and re-ported which of the features extracted by Wolberg et al. [54] were most significantfor classification.

In 2002, Estevez et al. [66] introduced a different approach for classificationbased on the Fuzzy Finite State Machine, but their system performed rather poorlygiving 19.4% error for the testing set of images. To extract features, the authorsfirst manually segment nuclei from the image and then apply a low-pass filter andin the following step topological map of a nuclei is created. The extracted featuresare texture based. Motivation for them was that benign cell textures have biggerhomogenous gray areas and more concentric contours than malignant cell textures.

Bagui et al. [67] recently introduced a classification algorithm applied to WBCD.The authors described a generalization of the rank nearest neighbor rule and ob-tained results that show a 97% recognition rate, which, according the authors, isbetter than that previously reported in the literature. From the above discussion wecan deduct that majority of work in the field of breast cancer detection and classifica-tion was performed by Street et al. and Wolberg et al. We can find other approachessuch as wavelet based approach of Weyn et al. [68]. Here the authors introduce atextural approach for chromatin description and claim that it has a 100% recognitionrate.

Another approach is one introduced by Schnorrenberg et al. [69] that uses re-ceptive fields for nuclei localization as an integral part of a bigger system, calledBASS. In 1996, they introduced a contentbased approach [70] and provided anextensive survey on existing histopathological systems [71]. The authors presentedtwo types of color-based features, luminance-based local features and global fea-tures. Luminance features were obtained from image RGB values. Global featuresare the variance and average of luminance in the image. They also introduce onetexture measure that is calculated according to the luminance variance and currentnucleus luminance. Approaches presented by Schnorrenberg et al. are mostly basedon histological samples rather than cytological. In 2000, they presented a descriptionof features used in their research [72] on classification of cryostat samples duringintra-operative examination based on feed-forward neural networks achieving thehighest accuracy of 76% on their own database.

In the literature we can also find some other approaches that involve segmentationof a breast cancer nuclei rather than classification. In 1996, Belhomme et al. [73]proposed a watershed based algorithm for segmentation of breast cancer cytologicaland histological images.Their algorithm is a more general version of the methoddescribed by Adams and Bischof [74]. The generalization involves the usage ofnumerous merging criteria. Authors use the segmentation principles described byBeucher in his PhD thesis [75]. This involves the decomposition of the segmentation


procedure into two steps. In the first step, the image is simplified based on a set ofmarkers. The second stage involves region decomposition by the construction of thewatershed lines [73]. The algorithm proposed by Belhomme et al. is the extensionof the Beucher and Meyer [76] method by introduction of a general segmentationoperator.

In 1998, Olivier et al. [77] introduced another extension to the watershed algo-rithm in addition to that of Belhomme et al. Their extension incorporates the colorinformation in the image regardless of the color space. The authors compared theirsegmentation results against the segmentation performed by three experts and theyreported the correctness of their method to be between 89.2% and 98.3% for thenuclei.

Another approach to nuclear segmentation is based on fuzzy cmeans clusteringand multiple active contours models described by Schpp et al. [78]. The authorsdescribe a level set active contours method, where the initial level set is obtained bythe fuzzy cmeans algorithm.

9 ComputerAided Breast Cancer Grading

In the previous section we described different approaches for breast cancer diagno-sis. Most of those systems discriminate only benign and malignant cases. For gooddiagnosis it is crucial to evaluate the malignancy grade. In cytology, the malignancyis graded according to the BloomRichardson scheme [79]. This system is basedon grading of cells polymorphy, the ability to reform histoformative structures, andmitotic index. All of these features are described by the Bloom-Richardson schemeas three factors that use a point based scale for assessing each feature. The malig-nancy of the tumor is assigned a grade that depends on the quantitative values of theabove factors and is determined by the summation of all awarded points for eachfactor. Depending on the value, the tumor is assigned with low, intermediate or highmalignancy grade.

In [80] we can see attempts at prognostication along with nuclear classification.For their grading approach, the authors used only nuclear features of a cell, whichcorrespond to the second factor in Bloom-Richardson grading scheme. They wereestimating the prognosis of the breast cancer according to these features. Furtherattempts for malignancy grading include VLSI approach introduced by Cheng etal. [81] in 1991 and applied in 1998 to breast cancer diagnosis [82]. In this method,the authors propose a parallel approach to tubule grading for histological slides.The authors divided their algorithm into four stages. The first stage consists of im-age enhancement for which purpose they use median filtering to remove artifacts. Instage two, the authors locate possible tubule formations by image thresholding witha threshold level known a priori. The next stage is a classification stage, where re-gions are classified as tubular formations. The features used in this study consists ofbrightness, bright homogeneity, circularity, size, and boundary colors. In the fourthstage, the authors count the number of tubular formations. The work presented bythe authors not only deals with histology but also only mentions grading using only


one factor on the Bloom-Richardson scale. The authors showed time improvementof the parallel algorithm that grades tubules to O(n) time while previously reportedrun time complexities were O(n2), where n is the size of the input data. In 1991,MacAulay et al. [83] introduced a graphics package for Bloom-Richardson gradingof histological tissue. Their application acts as a typical graphics program that al-lows user to pick the nuclei from the image and perform some basic calculations.This process is almost completely user dependent. The authors provide an extensivedescription of the interface of the package but no further information on computa-tion grading was found. Another approach found in literature is an algorithm basedon wavelet texture description of chromatin [68]. This work was also performed onhistological slides. The features calculated by the authors are calculated accordingto wavelet parameters and are divided into three groups. The first group are co-occurrence parameters that describe the color intensity in the image. The second setof parameters are densitometric parameters that are based on intensity values of thenucleus. The third group consists morphometric parameters that describe the geom-etry of the nucleus. Authors performed tests on their data-base of 83 histologicalslides and claim to have 100% classification rate. Such a high rate suggests a goodseparation between the classes.

In 2004, Gurevich and Murashov [84] proposed a method for chromatin struc-ture analysis based on scalespace approach of Florack and Kuijper [85]. The au-thors claim that chromatin distribution corresponds to the grade of malignancy. Thisstatement is supported by additional studies of Rodenacker [86, 87, 88] and Weyn etal. [89]. The authors also mention another approach to chromatin description. Thismethod uses heterogeneity, clumpiness, margination and radius of particles and wasintroduced by Young et al. [90]. The algorithm of Guverich and Murashov usestopological properties of isointensity manifolds in the spatial extrema neighbor-hoods [84]. Their algorithm is able to measure the number of chromatin particles inthe input image. For testing purposes the authors trained several classifiers achiev-ing a classification rate between 72% and 85.4%. In 2006, Gurevich et al. [91]described a system for automatic analysis of cytological slides for the lymphaticsystem tumors. The authors used a Gaussian filter for segmentation of a nuclei fromthe previously extracted blue channel of the image. The feature extraction part of theproposed system is the same as in [84] plus an additional 47 features described byChurakova et al. [92]. These features include a well known and widely used morpho-logical features such as the area of a nuclei, histogram features and features basedon a Fourier spectrum of a nucleus [91]. In this paper, the same choice of classifierswas used as in [84] but the accuracy increased and is claimed by the authors to beabove 90%. The authors did not provide an accurate error rate of their experimentsand therefore it is difficult to assess the accuracy of the proposed system.

To the best of our knowledge, currently there is no publicly available database andmost of the approaches presented in the literature are tested on the databases createdby the authors, which makes the comparison of the obtained classification resultswith those reported in the literature difficult. The only commonly used databasethat we came across during this study is the Wisconsin Breast Cancer Database,which was described earlier in this thesis. This database is freely available from the


authors web page [54]. In this study, some of the proposed features are the sameas in WBCD but the testing of the presented system on that database would belimited only to the classification stage due to the fact that WBCD is a database ofpre-extracted features.

In 2005 a commercial system for automated histopathological tissue grading wasreleased by QinetiQ [93]. According to the specifications and discussion with apathologist, the results obtained by this system seem to be difficult to confirm. Ac-cording to the authors, their system showed performance similar to the pathologistsduring clinical evaluation that was performed on 100 patients.

The most recent development in the field of automated breast cancer grading wasdescribed by Jelen in his PhD thesis [94]. There are also other recent approaches byNaik et al. [95] and Jelen et al. [96, 97, 98, 99, 100].

In [95] describe various segmentation methods such as level sets for classificationof prostate and breast cancer histological slides. The described system was able todistinguish between low and high malignancy grades with 80.52% accuracy whenautomatic classification was used. The accuracy described by Jelen in [94] was ashigh as 86.75% for cytological slides. The author in his thesis did an extensive studyof the features and classification methods to determine a set of features and the clas-sification method that will be able to classify the breast cancer malignancy intointermediate and high malignancy grades. Author also introduced a set of three newfeatures that are used for the determination of the first factor of BloomRichardsonscheme. These features where described in [96] and their discriminatory power weredescribed in [98]. Features that were introduced by Jelen include the area of groupedcells in the FNA slide (see Fig. 6), the number of groups that are visible on theslide and the third feature is a dispersion that describes if the cells in the image aregrouped or dispersed. Beside a set of so called low magnification features authorproposed the usage of 31 features that represented the nuclear structures of the cell.These features related to the second and third factor of the BloomRichardson grad-ing scheme. In the thesis, the author performed a set of classification tests performedthe calculations of the discriminatory power of the features to propose a set of fea-tures that are not correlated and provide the best classification results. From all ofthe tests, the author showed that the multilayer perceptron was the best performingclassifier. The 34 element feature vector was reduced to 15 features. Fig. 7 showsgraphically the correlation between the original set of 34 features. The features withthe best discriminatory power were the three low magnification features describedearlier and 12 nuclear features such as perimeter of a nucleus, convexity, xcentroidof the cell, nuclei orientation, its vertical projection, the _3 momentum feature,histogram mean, energy, textural homogeneity, red channel histogram mean, skewand width.

In [97] the authors did a comparative study of the discriminatory power of thelow magnification features against the features based on the cell nucleus. From theirstudy, one can notice that on average low magnification features perform better butthe best classification was recorded for a feature vector that consisted of both typesof features. In [99] the authors showed that the best classification was achievedfor the multilayer perceptron when the fuzzy cmeans segmentation was used. On


Fig. 6 FNA images: a) 100 resolution; b) 400 resolution.

Fig. 7 Correlation between extracted features: a) with variables in original order; b) withvariables regrouped by similarity.

average, for most tested classifiers, the best classifications where obtained when thelevel set segmentation was used.

The described work done by Jelen et al. was applied to a classification systembuilt and currently being tested in the pathological laboratory. In [100] the authorsshow that preliminary medical tests provide promising results and the automatedbreast cancer grading system performs with a high accuracy when applied to thereal and unseen data. The achieved accuracy was 81.96%.

10 Challenges and Future Developments

There are many challenging problems in automatic processing of histopathologies.The main problems include large variation of blood and cancer cells, occlusions,segmentation, low quality of images and difficulties in getting real data. We believethat these difficulties will be overcome with time.


References

[1] Newland, J.: The peripheral blood smear. In: Goldman, L., Ausiello, D. (eds.) CecilMedicine V, ch. 161, 23rd edn., Saunders Elsevier, Philadelphia (2007)

[2] Agby, G.: Leukopenia and leukocytosis. In: Goldman, L., Ausiello, D. (eds.) CecilMedicine, ch. 173, 23rd edn., Saunders Elsevier, Philadelphia (2007)

[3] Ramoser, H., Laurain, V., Bischof, H., Ecker, R.: Leukocyte segmentation and clas-sification in blood-smear images. In: 27th IEEE Annual Conference Engineering inMedicine and Biology, Shanghai, China, September 1-4 (2005)

[4] Al-Muhairy, J., Al-Assaf, Y.: Automatic white blood cell segmentation based on imageprocessing. In: 16th IFAC World Congress (2005)

[5] Prasad, B., Prasanna, S.M. (eds.): Speech, Audio, Image and Biomedical SignalProcessing using Neural Networks. Studies in Computational Intelligence, vol. 83.Springer, Heidelberg (2008); ISBN 978-3-540-75397-1

[6] Piuri, V., Scotti, F.: Morphological classification of blood leucocytes by microscopeimages. In: IEEE International Conference on Computational Intelligence Far Mea-surement Systems and Applications, Boston, MA, July14-16 (2004)

[7] Bentley, S., Lewis, S.: The use of an image analyzing computer for the quantificationof red cell morphological characteristics. British Journal of Hematology 29, 8188(1975)

[8] Rowan, R.: Automated examination of the peripheral blood smear. In: Automationand quality assurance in hematology, ch. 5, pp. 129177. Blackwell Scientific, Oxford(1986)

[9] Costin, H., Rotariu, C., Zbancioc, M., Costin, M., Hanganu, E.: Fuzzy rule-aided deci-sion support for blood cell recognition. Fuzzy Systems & Artificial Intelligence 7(1-3),6170 (2001)

[10] Albertini, M., Teodori, L., Piatti, E., Piacentini, M., Accorsi, A., Rocchi, M.: Auto-mated analysis of morphometric parameters for accurate definition of erythrocyte cellshape. Cytometry Part A 52A(1), 1218 (2003)

[11] Robinson, R., Benjamin, L., Cosgri, J., Cox, C., Lapets, O., Rowley, P., Yatco, E.,Wheeless, L.: Textural differences between aa and ss blood specimens as detected byimage analysis. Cytometry 17(2), 167172 (1994)

[12] Gering, E., Atkinson, C.: A rapid method for counting nucleated erythrocytes onstained blood smears by digital image analysis. Journal of Parasitology 90(4), 879881 (2004)

[13] Martelli, A.: An application of heuristic search methods to edge and contour detection.Communications of the ACM 19(2), 7383 (1976)

[14] Fleagle, S., Johnson, M., Wilbricht, C., Skorton, D., Wilson, R., White, C., Marcus, M.,Collins, S.: Automated analysis of coronary arterial morphology in cineangiograms:geometric and physiologic validation in humans. IEEE Transactions on Medical Imag-ing 8(4), 387400 (1989)

[15] Fleagle, S., Thedens, D., Ehrhardt, J., Scholz, T., Skorton, D.: Automated identificationof left ventricular borders from spin-echo magnetic resonance images. InvestigativeRadiology 26(4), 295303 (1991)

[16] Gonzalez, R., Woods, R.: Digital Image Processing (3rd Economy Edition). PrenticeHall, Englewood Cliffs (2008)

[17] Adjouadi, M., Fernandez, N.: An orientation-independent imaging technique for theclassification of blood cells. Particle & Particle Systems Characterization 18(2), 9198(2001)


[18] Di Ruberto, C., Dempster, A., Khan, S., Jarra, B.: Segmentation of blood images usingmorphological operators. In: 15th International Conference on Pattern Recognition(ICPR 2000), Barcelona, Spain, pp. 397400. IEEE, Los Alamitos (2000)

[19] Gauch, J.: Image segmentation and analysis via multiscale gradient watershed hierar-chies. IEEE Transactions on Image Processing 8(1), 6979 (1999)

[20] Kass, M., Witkin, A., Terzopoulos, D.: Snakes: Active contour models. InternationalJournal of Computer Vision 4, 321331 (1988)

[21] Ongun, G., Halici, U., Leblebicioglu, K., Atalay, V., Beksac, S., Beksac, M.: Auto-mated contour detection in blood cell images by an efficient snake algorithm. Nonlin-ear Analysis-Theory Methods & Applications 47(9), 58395847 (2001)

[22] Wang, X., He, L., Wee, W.G.: Deformable contour method: A constrained optimizationapproach. International Journal of Computer Vision 59(1), 87108 (2004)

[23] Ongun, G., Halici, U., Leblebicioglu, K., Atalay, V.V., Beksac, M., Beksac, S.: Featureextraction and classification of blood cells for an automated differential blood countsystem. In: Proc. IJCNN, vol. 4, pp. 24612466 (2001)

[24] Leznray, O., Elmoataz, A., Cardot, H., Gougeon, G., Lecluse, M., Elie, H., Revenu,M.: Segmentation of cytological images using color and mathematical morphology.European Congress of Stereology 18(1), 114 (1999)

[25] Ravi, B., Kumar, Danny, K., Joseph, Sreenivas, T.V.: Teager energy based blood cellsegmentation. In: International Conference on Digital Signal Processing, vol. 2, pp.619622 (2002)

[26] Ongun, G., Halici, U., Leblebicioglu, K., Atalay, V., Beksac, M., Beksak, S.: An auto-mated differential blood count system. In: IEEE Int. Conf. on Engineering in Medicineand Biology Society, vol. 3, pp. 25832586 (2001)

[27] Sinha, N., Ramakrishnan, A.: Automation of differential blood count. In: Conf. onConvergent Technologies for Asia-Pacific Region, vol. 2, pp. 547551 (2003)

[28] Comaniciu, D., Meer, P.: Cell image segmentation for diagnostic pathology. In: Suri,J., Setarehdan, S., Singh, S. (eds.) Advanced Algorithmic Approaches to Medical Im-age Segmentation: state-of-the-art application in cardiology, neurology, mammogra-phy and pathology, pp. 541558. Springer, Heidelberg (2001)

[29] Jiang, K., Liao, Q.-M., Dai, S.-Y.: A novel white blood cell segmentation scheme usingscale-space filtering and watershed clustering. In: Int. Conf. on Machine Learning andCybernetics, vol. 5, pp. 28202825 (2003)

[30] Hamghalam, M., Motameni, M., Kelishomi, A.E.: Leukocyte segmentation in giemsa-stained image of peripheral blood smears based on active contour. In: InternationalConference on Signal Processing Systems, pp. 103106. IEEE Computer Society, LosAlamitos (2009)

[31] Fodor, I., Kamath, C.: On denoising images using wavelet-based statistical techniques.Tech. rep., Lawrence Livermore National Laboratory, uCRL JC-142357 (2001)

[32] Sendur, L., Selesnick, I.: Bivariate shrinkage functions for wavelet-based denoising ex-ploiting interscale dependency. IEEE Transactions on Signal Processing 50(11), 27442756 (2002)

[33] Sendur, L., Selesnick, I.: A bivariate shrinkage function for wavelet-based denois-ing. In: IEEE International Conference on Acoustics, Speech and Signal Processing(ICASSP), vol. 2, pp. 12611264 (2002)

[34] Lim, J.S.: Two-Dimensional Signal and Image Processing. Prentice-Hall, EnglewoodCliffs (1990)


[35] Hong, V., Palus, H., Paulus, D.: Edge preserving filters on color images. In: Bubak, M.,van Albada, G.D., Sloot, P.M.A., Dongarra, J. (eds.) ICCS 2004. LNCS, vol. 3039, pp.3440. Springer, Heidelberg (2004)

[36] Babic, Z., Mandic, D.: An efficient noise removal and edge preserving convolutionfilter. In: 6th International Conference on Telecommunications in Modern Satellite,Cable and Broadcasting Service, vol. 2, pp. 538541 (2003)

[37] Nikolaou, N., Papamarkos, N.: Color reduction for complex document images. Inter-national Journal of Imaging Systems and Technology 19(1), 1426 (2009)

[38] Tomasi, C., Manduchi, R.: Bilateral filtering for gray and color images. In: Sixth In-ternational Conference on Computer Vision, January 4-7, pp. 839846 (1998)

[39] Papari, G., Petkov, N., Campisi, P.: Artistic edge and corner enhancing smoothing.IEEE Transactions on Image Processing 16(10), 24492462 (2007)

[40] Ntogas, N., Veintzas, D.: A binarization algorithm for historical manuscripts. In: 12thWSEAS International Conference on Communications, pp. 4151 (2008)

[41] Sauvola, J., Pietikainen, M.: Adaptive document image binarization. The Journal ofthe Pattern Recognition Society 33(2), 225236 (2000)

[42] Otsu, N.: A threshold selection method from gray-level histograms. IEEE Transactionson System, Man and Cybernetics 9(1), 6266 (1979)

[43] Habibzadeh, M., Krzyzak, A., Fevens, T., Sadr, A.: Counting of RBCs and WBCs innoisy normal blood smear microscopic images. In: SPIE Medical Imaging (February2011)

[44] Vincent, L.: Fast opening functions and morphological granulometries. Image Algebraand Morphological Image Processing 2300, 253267 (1994)

[45] Lin, Y.-C., Tsai, Y.-P., Hung, Y.-P., Shih, Z.-C.: Comparison between immersion-basedand toboggan-based watershed image segmentation. IEEE Transactions on Image Pro-cessing 15(3), 632640 (2006)

[46] Duda, R., Hart, P., Stork, D.: Pattern Classification, 2nd edn. Wiley Interscience Pub-lishers, Hoboken (2000)

[47] Cheng, H., Shi, X., Min, R., Cai, X., Du, H.N.: Approaches for Automated Detectionand Classification of Masses in Mammograms. Pattern Recognition 39(4), 646668(2006)

[48] Bottema, M., Slavotinek, J.: Detection and Classification of Lobular and DCIS (smallcell) Microcalcifications in Digital Mammograms. Pattern Recognition Letters 21(13-14), 12091214 (2000)

[49] Cheng, H., Cui, M.: Mass Lesion Detection with a Fuzzy Neural Network. PatternRecognition 37, 11891200 (2004)

[50] Cheng, H., Wang, J., Shi, X.: Microcalcification Detection using Fuzzy Logic andScale Space Approaches. Pattern Recognition 37(2), 363375 (2004)

[51] De Santo, M., Molinara, M., Tortorella, F., Vento, M.: Automatic Classification ofClustered Microcalcifications by a Multiple Expert System. Pattern Recognition 36(7),14671477 (2003)

[52] Grohman, W., Dhawan, A.: Fuzzy Convex Set-based Pattern Classification for Anal-ysis of Mammographic Microcalcifications. Pattern Recognition 34(7), 14691482(2001)

[53] Zhang, P., Verma, B., Kumar, K.: Neural vs. Statistical Classifier in Conjunction withGenetic Algorithm Based Feature Selection. Pattern Recognition Letters 26(7), 909919 (2005)


[54] Wolberg, W., Mangasarian, O.: Multisurface Method of Pattern Separation for MedicalDiagnosis Applied to Breast Cytology. Proceedings of National Academy of Science,USA 87, 91939196 (1990)

[55] Mangasarian, O., Setiono, R., Wolberg, W.: Pattern Recognition via Linear Program-ming: Theory and Application to Medical Diagnosis. In: Large-Scale Num. Opt., pp.2231. SIAM, Philadelphia (1990)

[56] Street, W.N., Wolberg, W.H., Mangasarian, O.L.: Nuclear Feature Extraction forBreast Tumor Diagnosis. In: Imaging Science and Technology/Society of Photo-graphic Instrumentation Engineers 1993 International Symposium on Electronic Imag-ing: Science and Technology, San Jose, California, vol. 1905, pp. 861870 (1993)

[57] Street, N.: Cancer Diagnosis and Prognosis via Linear-Programming-Based MachineLearning. Ph.D. thesis, University of Wisconsin (1994)

[58] Street, N.: Xcyt: A System for Remote Cytological Diagnosis and Prognosis of BreastCancer. In: Jain, L. (ed.) Soft Computing Techniques in Breast Cancer Prognosis andDiagnosis, pp. 297322. World Scientific Publishing, Singapore (2000)

[59] Lee, K., Street, W.: A Fast and Robust Approach for Automated Segmentation ofBreast Cancer Nuclei. In: Proceedings of the Second IASTED International Confer-ence on Computer Graphics and Imaging, Palm Springs, CA, pp. 4247 (1999)

[60] Lee, K., Street, W.: Model-based Detection, Segmentation and Classification for ImageAnalysis using On-line Shape Learning. Machine Vision and Applications 13(4), 222233 (2003)

[61] Wolberg, W.H., Street, W.N., Mangasarian, O.L.: Breast Cytology Diagnosis Via Dig-ital Image Analysis. Analytical and Quantitative Cytology and Histology 15, 396404(1993)

[62] Wolberg, W.H., Street, W.N., Mangasarian, O.L.: Machine Learning Techniques toDiagnose Breast Cancer from Image-Processed Nuclear Features of Fine Needle As-pirates. Cancer Letters 77, 163171 (1994)

[63] Walker, H.J., Albertelli, L., Titkov, Y., Kaltsatis, P., Seburyano, G.: Evolution of NeuralNetworks for the Detection of Breast Cancer. In: Proceedings of International JointSymposia on Intelligence and Systems, pp. 3440 (1998)

[64] Walker, H.J., Albertelli, L.: Breast Cancer Screening Using Evolved Neural Networks.In: IEEE International Conference on Systems, Man and Cybernetics, vol. 2, pp. 16191624 (1998)

[65] Nezafat, R., Tabesh, A., Akhavan, S., Lucas, C., Zia, M.: Feature Selection and Clas-sification for Diagnosing Breast Cancer. In: Proceedings of International Associationof Science and Technology for Development International Conference, pp. 310313(1998)

[66] Estevez, J., Alayon, S., Moreno, L.: Cytological Breast Cancer Fine Needle AspirateImages Analysis with a Genetic Fuzzy Finite State Machine. In: Conference Board ofthe Mathematical Sciences, CBMS 2002, pp. 2126 (2002)

[67] Bagui, S., Bagui, S., Pal, K., Pal, N.: Breast Cancer Detection using Rank NearestNeighbor Classification Rules. Pattern Recognition 36(1), 2534 (2003)

[68] Weyn, B., van de Wouwer, G., van Daele, A., Scheunders, P., van Dyck, D., van Marck,E., Jackob, W.: Automated Breast Tumor Diagnisis and Grading Based on WaveletChromatin Texture Description. Cytometry 33, 3240 (1998)

[69] Schnorrenberg, F., Pattichis, C., Kyriacou, K., Schizas, C.: Detection of Cell Nuclei inBreast Cancer Biopsies using Receptive Fields. In: IEEE Proceedings of Engineeringin Medicine and Biology Society, pp. 649650 (1994)


[70] Schnorrenberg, F., Pattichis, C., Kyriacou, K., Schizas, C.: Content-based Descriptionof Breast Cancer Biopsy Slides. In: Proc. Intl. EuroPACS Mtg., pp. 136140 (1996)

[71] Schnorrenberg, F., Pattichis, C., Kyriacou, K., Vassiliou, M., Schizas, C.: Computer-aided Classification of Breast Cancer Nuclei. Technology & Health Care 4(2), 147161 (1996)

[72] Schnorrenberg, F., Tsapatsoulis, N., Pattichis, C., Schizas, C., Kollias, S., Vassiliou,M., Adamou, A., Kyriacou, K.: A modular neural network system for the analysisof nuclei in histopathological sections. IEEE Engineering in Medicine and BiologyMagazine 19, 4863 (2000)

[73] Belhomme, P., Elmoataz, A., Herlin, P., Bloyet, D.: Generalized region growing op-erator with optimal scanning: application to segmentation of breast cancer images.Journal of Microscopy 186, 4150 (1997)

[74] Adams, R., Bischof, L.: Seeded region growing. IEEE Transactions on Pattern Analy-sis and Machine Intelligence 16, 641647 (1994)

[75] Beucher, S.: Segmentation dimages et morphologie mathmatique. Ph.D. thesis,Ecole National Suprieur des Mines de Paris (1990)

[76] Beucher, S., Meyer, F.: Mathematical Morphology in Image Processing, ch. 12, pp.433481. Marcel Dekker, New York (1992)

[77] Lezoray, O., Elmoataz, A., Cardot, H., Gougeon, G., Lecluse, M., Revenu, M.: Seg-mentation of cytological images using color and mathematical morphology. In: Euro-pean Conference on Stereology, Amsterdam, Netherlands, p. 52 (1998)

[78] Schpp, S., Elmoataz, A., Fadili, J., Herlin, P., Bloyet, D.: Image segmentation viamultiple active contour models and fuzzy clustering with biomedical applications. In:The 15th International Conference on Pattern Recognition, ICPR 2000, Barcelona,Spain, vol. 1, pp. 622625 (2000)

[79] Bloom, H., Richardson, W.: Histological Grading and Prognosis in Breast Cancer.British Journal of Cancer 11, 359377 (1957)

[80] Mangasarian, O., Street, W., Wolberg, W.: Breast Cancer Diagnosis and Prognosis viaLinear Programming. Operations Research 43(4), 570576 (1994)

[81] Cheng, H., Li, X., Riodan, D. S., Scrimger, J.N.: A Parallel Approach to Tubule Grad-ing in Breast Cancer Lesions and its VLSI Implementation. In: Fourth Annual IEEESymposium on Computer-Based Medical Systems, pp. 322329 (1991)

[82] Cheng, H., Wu, C., Hung, D.: VLSI for Moment Computation and its Application toBreast Cancer Detection. Pattern Recognition 31(8), 13911406 (1998)

[83] MacAulay, M., Scrimger, J., Riodan, D., Cheng, H.: An Interactive Graphics Packagewith Standard Examples of the Bloom and Richardson Histological Grading Tech-nique. In: Fourth Annual IEEE Symposium on Computer-Based Medical Systems, pp.108112 (1991)

[84] Gurevich, I., Murashov, D.: Method for early diagnostics of lymphatic system tumorson the basis of the analysis of chromatin constitution in cell nucleus images. In: The17th International Conference on Pattern Recognition, ICPR 2004, Cambridge, UK,pp. 806809 (2004)

[85] Florack, L., Kuijper, A.: The topological structure of scalespace images. Journal ofMathematical Imaging and Vision 12(1), 6580 (2000)

[86] Rodenacker, K.: Applications of topology for evaluating pictorial structures. In: Theo-retical Foundations of Computer Vision, pp. 3546. AkademieVerlag, Berlin (1993)

[87] Rodenacker, K.: Quantitative microscope image analysis for improved diagnosis andprognosis of tumors in pathology. In: Creaso Info Medical Imaging, Creaso GmbH,Gilching, vol. 22 (1995)


[88] Rodenacker, K., Bengtsson, E.: A feature set for cytometry on digitized microscopicimages. Anal. Cell Pathol. 25(1), 136 (2003)

[89] Weyn, B., Van de Wouwer, G., Koprowski, M., et al.: Value of morphometry, textureanalysis, densitometry and histometry in the differential diagnosis and prognosis ofmalignant mesthelioma. Journal of Pathology 4(189), 581589 (1999)

[90] Young, I., Verbeek, P., Mayall, B.: Characterization of chromatin distribution in cellnuclei. Cytometry 7(5), 467474 (1986)

[91] Gurevich, I., Kharazishvili, D., Murashov, D., Salvetti, O., Vorobjev, I.: Technologyfor automated morphologic analysis of cytological slides. methods and results. In:The 18th International Conference on Pattern Recognition (ICPR 2006), Hong Kong,China, pp. 711714 (2006)

[92] Churakova, Z., Gurevich, I., Jernova, I., et al.: Selection of diagnostically valuable fea-tures for morphological analysis of blood cells. Pattern Recognition and Image Anal-ysis: Advances in Mathematical Theory and Applications 13(2), 381383 (2003)

[93] Qineti, Q.: Automated Histopathology Breast Cancer Analysis and Diagnisis Sys-tem.Data Sheet (2005), http://www.qinetiq.com/

[94] Jelen, .: Computerized Cancer Malignancy Grading of Fine Needle Aspirates. Ph.D.thesis, Concordia University (2009)

[95] Naik, S., Doyle, S., Agner, S., Madabhushi, A., Feldman, M., Tomaszewski, J.: Au-tomated gland and nuclei segmentation for grading of prostate and breast cancerhistopathology. In: Proceedings of the IEEE International Symposium on BiomedicalImaging, pp. 284287 (2008)

[96] Jelen, ., Fevens, T., Krzyzak, A.: Classification of breast cancer malignancy usingcytological images of fine needle aspiration biopsies. Int. J. Math. Comput. Sci. 18,7583 (2008)

[97] Jelen, ., Krzyzak, A., Fevens, T.: Comparison of pleomorphic and structural featuresused for breast cancer malignancy classification. In: Bergler, S. (ed.) Canadian AI.LNCS (LNAI), vol. 5032, pp. 138149. Springer, Heidelberg (2008)

[98] Jelen, ., Fevens, T., Krzyzak, A., Jelen, M.: Discriminatory power of cells groupingfeatures for breast cancer malignancy classification. In: Proceedings of the Interna-tional Federation for Medical and Biological Engineering, vol. 21(3), pp. 559562.Springer, Heidelberg (2008)

[99] Jelen, ., Fevens, T., Krzyzak, A.: Influence of nuclei segmentation on breast cancermalignancy classification. In: Proceedings of SPIE, vol. 7260, pp. 7260147260149(2009)

[100] Jelen, ., Lipinski, A., Detyna, J., Jelen, M.: Clinical verification of computerizedbreast cancer malignancy grading. BioAlgorithms and MedSystems 6(12), suppl. 1,8182 (2010)

Application of Pattern Recognition Techniques for the Analysis of Histopathological ImagesIntroductionManual Analysis of Blood Smear ImagesAutomatic Processing of Blood Smear ImagesMethodology and AlgorithmsImage Acquisition and DenoisingEdge PreservationBinarizationSize EstimationSegmentation

Conclusion: Segmentation of Microscopic ImageryCancer Cells GradingBreast Cancer DiagnosisComputerAided Breast Cancer DiagnosisComputerAided Breast Cancer GradingChallenges and Future DevelopmentsReferences

Documents

Application of Pattern Recognition Techniques for the Analysis of Histopathological Images-libre