You are on page 1of 18


7, 2014


Methods for Nuclei Detection, Segmentation,
and Classification in Digital Histopathology:
A Review—Current Status and Future Potential
Humayun Irshad, Student Member, IEEE, Antoine Veillard, Ludovic Roux, and Daniel Racoceanu, Member, IEEE
Methodological Review

Abstract—Digital pathology represents one of the major evolutions in modern medicine. Pathological examinations constitute the
gold standard in many medical protocols, and also play a critical
and legal role in the diagnosis process. In the conventional cancer diagnosis, pathologists analyze biopsies to make diagnostic and
prognostic assessments, mainly based on the cell morphology and
architecture distribution. Recently, computerized methods have
been rapidly evolving in the area of digital pathology, with growing
applications related to nuclei detection, segmentation, and classification. In cancer research, these approaches have played, and
will continue to play a key (often bottleneck) role in minimizing
human intervention, consolidating pertinent second opinions, and
providing traceable clinical information. Pathological studies have
been conducted for numerous cancer detection and grading applications, including brain, breast, cervix, lung, and prostate cancer
grading. Our study presents, discusses, and extracts the major
trends from an exhaustive overview of various nuclei detection,
segmentation, feature computation, and classification techniques
used in histopathology imagery, specifically in hematoxylin–eosin
and immunohistochemical staining protocols. This study also enables us to measure the challenges that remain, in order to reach
robust analysis of whole slide images, essential high content imaging with diagnostic biomarkers and prognosis support in digital
Index Terms—Digital pathology, histopathology, microscopic
analysis, nuclei classification, nuclei detection, nuclei segmentation, nuclei separation.

ATHOLOGY is the microscopic study of the cell morphology supplemented with in situ molecular information. The
tissue sample is removed from the body and then prepared for
viewing under the microscope by placing it in a fixative, which
stabilizes the tissue to prevent decay. For the sake of visualiz-


Manuscript received October 1, 2013; revised December 7, 2013; accepted
December 16, 2013. Date of publication December 18, 2013; date of current
version April 28, 2014. This work was supported in part by the French National
Research Agency ANR, Project MICO, reference ANR-10-TECS-015.
H. Irshad and L. Roux are with the University Joseph Fourier, 38041 Grenoble, France, and Image Pervasive Access Lab, Centre national de la recherch´e
scientifique Unit Mixed International 2955, Singapore (e-mail: humayun.;
A. Veillard and D. Racoceanu are with the Sorbonne Universit´es, University
of Pierre and Marie Curie, Unit Mixed International 2955, 75005 Paris, France,
and also with the Image and Pervasive Access Laboratory, Singapore (e-mail:;
Digital Object Identifier 10.1109/RBME.2013.2295804

ing under the microscope, different components of the tissue
are dyed with different stains. Then, different staining techniques are applied to reveal specific tissue components under the
The pathologist plays a central role in therapeutic decision
making [1], [2]. Accordingly, diagnosis from pathology images
remains the “gold standard” in diagnosing a number of diseases
including most cancers [3]. Diagnosing a disease after manually analyzing numerous biopsy slides represents a labor intensive work for pathologists. Thanks to recent advances in digital
pathology, the automated recognition of pathology patterns in a
high-content whole slide image (WSI) has the potential to provide valuable assistance to the pathologist in his daily practice.
Researchers in pathology have been familiar with the importance of quantitative analysis of pathological images. Quantitative analysis can be used to support pathologists decision about
the presence or the absence of a disease, and also to help in
disease progression evaluation. In addition, quantitative characterization is important, not only for clinical usage (e.g., to
increase the diagnostic reliability), but also for research applications (e.g., drug discovery [4] and biological mechanisms
of disease [5]). As a consequence, the use of computer-aided
diagnosis (CAD) in pathology can substantially enhance the
efficiency and accuracy of pathologists decisions, and overall
benefit the patient.
Earliest works in the field date back to the early 90s [6]–[8]
but are in relatively small number, presumably due to the limited penetration of digital equipment in pathology. Thanks to
recent advances in digital pathology, numerous cancer detection
and grading applications have been proposed, including brain
[9]–[11], breast [12]–[21], cervix [22], [23], liver [24], lung [25],
and prostate [26]–[28] cancer grading.
Among the various studies, automated nuclei segmentation
and classification is a recurring task, particularly difficult on
pathology images. Indeed, the detection and segmentation of
nuclei in cytopathology images are generally facilitated due
to the well-separated nuclei and the absence of complicated
tissue structures. In contrast, the segmentation of nuclei on
histopathological images (tissue preserving its original structure) is more difficult since most of the nuclei are often part of
histological structures presenting complex and irregular visual
A review on automated cancer diagnosis based on histopathological images [29], [30] and a review on histopathological

1937-3333 © 2014 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution
requires IEEE permission. See standards/publications/rights/index.html for more information.

Nuclei are susceptible to exhibit a wide variety of patterns (related to the distribution of chromatin. and classification. H&E staining has been used by pathologists for over a hundred years [31] and is still widely used for observing morphological features of the tissue under white light microscopes. Section II introduces the different image modalities in histopathology. fast slide scanners are used to generate digital images that contain relevant information about the specimen at a microscopic level. and classification. pointing to future research directions and open problems related to nuclei detection. 2014 TABLE I DESCRIPTION OF NOTATION image analysis [5] already exist in the literature. segmentation. This paper is intended as a comprehensive state-ofthe-art survey on the particular issues of nuclei detection. VOL. They are capable of digitizing complete slides usually at ×20 or . 1(b) shows an example of IHC under light microscopy. A list of symbols and notation commonly used in this paper is shown in Table I. genetic progression. stroma. Examples of H&E and IHC images. By revealing the presence or absence of specific proteins in the observed tissue. IHC is a technique used for diagnosing whether a cancer is benign or malignant and for determining the stage of a tumor. the sections are dyed with one or more stains. segmentation. specific therapeutic treatments adapted for this type of cancer are selected. we highlight the challenges in nuclei detection. For tissue components visualization under a microscope. Fig. (b) IHC. and suggests ways to overcome them. Section V addresses the challenges in nuclei detection. Fig. This paper is organized in six sections. and classification methods restricted to two widely available types of image modalities: hematoxylin-eosin (H&E) and immunohistochemical (IHC). 1. (a) H&E. H&E staining is a widespread staining protocol in pathology. They embark one or multiple lenses to magnify the sample and capture digital images with a camera. 7. We conclude with a discussion.98 IEEE REVIEWS IN BIOMEDICAL ENGINEERING.) with a pink color [see Fig. According to the proteins revealed by IHC. 1(a)]. IHC helps in determining which cell type is at the origin of a tumor. while eosin stains other structures (cytoplasm. segmentation. addressing different types of problems associated with different image modalities. Section IV illustrates the recent advances in nuclei detection. and classification methods used in histopathology. II. After staining. In Section III. STAINING AND IMAGE MODALITIES IN DIGITAL PATHOLOGY Digital pathology is the microscopic investigation of a biopsy or surgical specimen that is chemically processed and sectioned onto glass slides to study cancer expression. Hematoxylin stains nuclei in dark blue color. and cellular morphology for cancer diagnosis and prognosis. prominent nucleolus) that are diagnostically significant. segmentation. etc. segmentation. and classification.

if I(i) ≥ T  (2) I (i) = 0. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY 99 Fig. 2(a)]. malignancy of the disease. Another thresholding technique is local (adaptive) thresholding that handles nonuniform illumination. referred to as nuclear pleomorphism. infiltration of LN in breast cancer is related to patient survival and outcome [33]. (b) Overlaps. SEGMENTATION. 3. [32] presented a robust Bayesian color segmentation algorithm that dynamically estimates the probability density functions describing the color and spatial properties of salient objects. (d) EN (Mitosis). In Section IV-A. Numerous works. Nuclei may look very different according to a number of factors such as nuclei type. appearance. mitotic nuclei (MN). we introduce the most commonly used image processing methods. where i ∈ U. In addition. and c is the number of channels (also called colors). To reduce these differences in illumination. (b) EN. use a single or a combination of these image processing methods for preprocessing. CHALLENGES IN NUCLEI SEGMENTATION AND CLASSIFICATION Among the different types of nuclei. m − 1]] × [[0. 2(b)]. 3}. IV. uniform light spectrum is used to highlight the tissue slide. 2(c)]. Furthermore. stored in a format that enables fast zooming and panning. The variation in nuclei shape. IV-C. III. sample thickness. then I  is a binary image of I as  1. To address the problem of color nonstandardness.) and image acquisition (artifacts introduced by the compression of the image. The output of the digital scanners is multilayered images.IRSHAD et al. (c) EN (Cancer). and IV-E. two types are usually the object of particular interest: lymphocyte and epithelial nuclei. Automated nuclei segmentation is now a well-studied topic for which a large number of methods have been described in the literature and new methodologies continue to be investigated. SEGMENTATION. The aspect of nuclei is critical for evaluating the existence of disease and its severity. usually c ∈ {1. and nuclei life cycle. A successful image processing approach will have to overcome these issues in a robust way in order to maintain a high level in the quality and accuracy in all situations. and texture during nuclei life cycle. IV-D. NUCLEI DETECTION. clusters and have heterogeneous aspects. 3) make the nuclei detection. All these problems (highlighted in Fig. segmentation. 1]c (1) where U = [[0. Lymphocyte nuclei (LN) are inflammatory nuclei having regular shape and smaller size than epithelial nuclei (EN) [see Fig. m and n are the number of rows and columns. n − 1]] are the pixels. A threshold value can be estimated using computational methods like the Otsu method which determines an optimal threshold by minimizing the intraclass variance [37]. segmentation. Examples of challenging nuclei to detect and segment. A. 2(d)]. most camera technologies have low response to short wavelength (blue) illumination and have a high sensitivity at long wavelength (red to infrared) regions. It can be determined by either splitting an image into subimages and calculating thresholds for . Detection. segmentation. Lymphocyte is a type of white blood cell that has a major role in the immune system. An image I is a function I : U −→ [0. A part of image I denoted Ij is a restriction of I to a connected subset of pixels. presence of digital noise. and clearly visible nucleoli as compared to normal EN [see Fig. For example. respectively. etc. nuclei pleomorphism has diagnostic value for cancer grading [34]–[36]. most slide scanners provide standard packages to normalize and correct spectral and spatial illumination variations. mitotic count is also an important prognostic parameter in breast cancer grading [34]. Furthermore. (a) Blur. For illumination. presence of foreign artifacts or damage to the tissue sample. If T is a global threshold. and separation. size. Threshold T can be global or local. etc. EN are larger in size. Similarly.). Image Processing Methods We begin with basic definitions. 1) Thresholding: Thresholding is a method used for converting intensity image I into a binary image I  by assigning all pixels to the value one or zero if their intensity is above or below some threshold T . Different types of nuclei. may have heterogeneous chromatin distribution. and classification of nuclei in routinely stained histopathological images pose a difficult computer vision problem due to high variability in images caused by a number of factors including differences in slide preparation (dyes concentration. Monaco et al. (c) Heterogeneity. In high-grade cancer tissue. The microscope setup. AND CLASSIFICATION Nuclei detection and segmentation are important steps in cancer diagnosis and grading. (a) LN. 2. otherwise. specific features of the slide scanner. detection. and staining may cause uneven illumination. described in Sections IV-B.: METHODS FOR NUCLEI DETECTION. ×40 magnifications. nuclei are often organized in overlapping Fig. I(i) is the ith pixel value in the image I. and classification a challenging problem. is another factor of complexity [see Fig. evenness of the cut. irregular boundaries. Nonpathological EN have nearly uniform chromatin distribution with smooth boundary [see Fig.

The erosion and dilation of the binary image I by the structuring element S ∈ Z × Z are defined as Erosion: Uf  S = {x|∀s ∈ S. x + s ∈ Uf } Dilation: Uf ⊕ S = {x + s|x ∈ I ∧ s ∈ S}. texture. The segmentation results in a partition of I into regions (I1 . n. widely used in image segmentation.. called catchment basin. j) j ∈N (i) where N (i) is neighbors of pixel i and d(i. predefined shape. are deformable splines that can be used to depict the contour of objects in an image using gradient information by seeking to minimize an energy function [43]. The general ACM is defined using the energy function E over the contour points c as (12) E = (αEInt (c) + βEIm g (c) + γEExt (c)) dc 3) Region Growing: Region growing [41] is an image segmentation method consisting of two steps. The basic mathematical definition contains lower slope LS(i) that is the maximum slope connecting pixel i in the image I to its neighbors of lower altitude as   I(i) − I(j) (9) LS(i) = max d(i. j). It allows to classify every point of a topographic surface as either belonging to the catchment basin associated with one of the local minima or to the watershed line. 2. ∀Ii . j) = LS(j) · d(i. drawing conclusions on how this shape fits or misses the shapes in the image. if I(i) = I(j). 5) Active Contour Models and Level sets: Active contour models (ACMs) or deformable models. color. The basic idea is to probe an image I with a simple. . The P r that are often used are gray level (average intensity and variance). In ).100 IEEE REVIEWS IN BIOMEDICAL ENGINEERING. if I(i) < I(j) ⎩1 (LS(i) + LS(j)) · d(i.. Let P r(Ii ) is a logical predicate which measures the similarity of a region Ii . This simple probe is called the structuring element and is a subset of the image. The typically used binary structuring elements are crosses. Details about watershed can be found in [42]. Let I : U −→ {0. . ik +1 ) (11) k =0 (8) where Π is the set of all paths from i to j. Two other major operations in morphology are opening ◦ and closing •. (3) The basic effect of erosion (dilation) operator on an image is to shrink (enlarge) the boundaries of foreground pixels. it fuses narrow breaks and fills small holes and gaps in the image. 4) Watershed: Watershed is a segmentation method that usually starts from specific pixels called markers and gradually floods the surrounding regions of markers. 7. The cost of moving from pixel i to j is defined as ⎧ if I(i) > I(j) ⎨ LS(i) · d(i. and open disks. . . it eliminates small objects and sharpens peaks in the object. . . . 1} be a binary image and Uf = I −1 ({1}) be the foreground pixels.i t )∈Π (5) t−1  d(ik . by treating pixel values as a local topography. Ij (i . The two basic morphological operators are the erosion  and the dilation ⊕. In case of i = j. Catchment basins are separated topographically from adjacent catchment basins by maximum altitude lines called watershed lines. ik +1 ) · cost (ik . 2) P r(Ii ∪ Ij ) = FALSE. Opening is mathematically defined as Uf ◦ S = [Uf  S] ⊕ S. the lower slope is forced to be zero. squares. The energy function is often defined to penalize discontinuity in the curve shape and graylevel discontinuity along the contour [12]. The watershed transformation is usually computed on the gradient image instead of the intensity image. Opening is an erosion of an image followed by dilation. Closing is mathematically defined as Uf • S = [Uf ⊕ S]  S.. j). I2 . and shape related.. . In case of nuclei segmentation. (4) Closing is a dilation of an image followed by an erosion. 2) Morphology: Morphology is a set-theoretic approach that considers an image as the elements of a set [39] and process images as geometrical shapes [40]. 2 (10) The topographical distance between the two pixels i and j is expressed as min (i 0 . 2014 each subimage or examining the image intensity in the pixel’s neighborhood [38]. cost(i. the contour points that yield the minimum energy level form the boundary of nuclei. j). j) is the Euclidean distance between pixels i and j. so that the following conditions hold: 1) P r(Ii ) = TRUE for all i = 1. VOL. The first step is the selection of seed points and the second step is a classification of neighboring pixels to determine whether those pixels should be added to the region or not by minimizing a cost function.

and EExt is the userdefined force or prior knowledge of object to control the contour (referring as external energy). called snakes. c . It is defined as G(I) = [Uf • S] − [Uf ◦ S]. The white top-hat transform is defined as the difference between image I and its opening as Tw (I) = Uf − [Uf ◦ S]. α. (7) In addition. (6) The black top-hat transform is defined as the difference between image I and its closing as Tb (I) = Uf − [Uf • S]. There are two main forms of ACMs. and γ are empirically derived constants. β. where EInt controls the shape and length of the contour (often called internal energy). is useful for edge detection. An explicit parametric representation of the contour. is robust to image noise and boundary gaps as it constrains the extracted White and black top-hat transforms are two other operations derived from morphology. which is the difference between the dilation and the erosion of a given image.= j) adjacent regions. morphological gradient. EIm g influences adjustment of local parts of the contour to the image values regardless of the contour geometry (referring as image energy). They allow to extract small elements and details from given images.

σk2 . In the following sections. in which an image is conceptualized as weighted undirected graph G(V. Preprocessing Preprocessing can be performed to compensate for adverse conditions such as the presence of batch effects. These image processing methods are extensively used in recently proposed frameworks for preprocessing. conditional probability P (Ck |I(i)) is computed as (t) (t) (t ) w N (I(i)|μk σk2 ) P (Ck |I(i))(t) = K k (t) . 2) Assign cluster label to each pixel in the image that minimizes the distance between the pixel and the cluster center. or other image parameters recurring across multiple images. region of interest (ROI) detection can also be performed in order to reduce the processing time. Ncut(A. (0) (0) (0) 1) Initialization: The parameters μk . and classification in histopathology as shown in Table II. The basic idea is to determine level curves from a potential function. 1) Pick K cluster centers. μk . B) = w(u. segmentation. the implicit ACM. E) by representing nodes V with pixels. The image is modeled according to the probability distribution P (I(i)) = K  wk N (I(i)|μk . σk2 . called level sets. 3) Recompute the cluster centers by averaging all the pixels in the cluster. Additionally. t) v ∈V B . σk2 ) (13) k =1 where K is the number of clusters (objects in the image). 8) Graph Cuts: Graph cuts (Gcuts) refers to a wide family of algorithms. the minimum cut criterion favors small isolated regions. and wk are randomly initialized for each cluster Ck . It is mathematically defined as cut(A. k. variance. anisotropy. but they are not robust to boundary gaps and have other deficiencies as well [44]. Its robustness depends mainly on the initialization of clusters. The basic procedure used to find the minimum Ncut is explained here [48]. snakes are restricted for topological adaptability of the model. However.IRSHAD et al. u ∈V A . This total weight is called a cut  cut(A.v ∈V B An intuitive way is to look for the minimum cut in the graph. Batch effect refers to unevenness in illumination. B) cut(A. v). either randomly or based on some heuristic.t∈V w(v. we compiled a list of existing frameworks for nuclei detection. separation. other features can be used like texture. nuclei detection. Carson et al. and location. The normalized cut (Ncut) solves this problem by computing the cut cost as a fraction of total edge connections to all the nodes in the graph. segmentation. and wk of each cluster Ck are now maximized using all pixels and the computed probabilities P (Ck |I)(t) from expectation step as U P (Ck |I(i))(t) · I(i) (t+1) μk = i U (15) (t) i P (Ck |I(i)) U (t+1) 2 P (Ck |I(i))(t) · (I(i) − μk ) (t+1) σk = i (16) U (t) i P (Ck |I(i)) U P (Ck |I(i))(t) (t+1) . B) + . B) = B. The basic algorithm is as follows. 2) Expectation: For each pixel I(i) and cluster Ck . The difference is typically based on the pixel value. we discuss how different image processing methods have been used. texture. t) (19) Ncut value would not be small for the cut that partitions isolating points. the total number of pixels in I. 7) Probabilistic Models: Probabilistic models can be viewed as an extension of K-means clustering. The Gcuts method partitions the graph into disjoint subgraphs so that similarity is high within the subgraphs and low across different subgraphs. The EM is based on the following four steps.t∈V w(u. EA ) and B(VB . or any features between two pixels. Alternatively. in case of splitting or merging of contours. Based on these image processing methods. weighted edges E with similarity (affinity) measure between nodes W : V 2 −→ R+ . and contrast. and classification. separation. (18) u ∈V A . Noise reduction and artifacts elimination can also be performed prior to detection and segmentation. 1) Illumination Normalization: The illumination can be corrected either by using white shading correction or by estimating .: METHODS FOR NUCLEI DETECTION. color. 4) Termination: Steps 2) and 3) are repeated until parameters converge. and contrast is a measure of homogeneity of pixels. [47] described the use of a new set of texture features polarity. The parameters of GMM are estimated from training data using the computation method like expectation maximization (EM) [46] that iteratively finds maximum likelihood. Gaussian mixture models (GMMs) are a popular parametric probabilistic model represented as weighted sum of Gaussian cluster densities. The degree of dissimilarity between two subgraphs A and B can be computed as the sum of weights of the edges that must be removed to separate A(VA . SEGMENTATION. (t) 2 ( t ) ) j =1 wj N (I(i)|μj σj (t) (t ) (14) (t) 3) Maximization: The parameters μk . and weight of cluster w The wk are positive real values such that K k =1 k = 1. which are not useful in finding large uniform regions. Instead of pixel values. and wk are mean. σk2 . A similarity measure is computed from intensity. spatial distribution. 4) Repeat steps 2) and 3) until convergence is attained or no pixel changes its cluster. EB ). because the cut value will be a large percentage of the total connection from that set to the others. is specifically designed to handle topological changes. However. anisotropy is a ratio of the eigenvalues of the second moment matrix. respectively. or a weighted combination of these factors. Polarity is a measure of a gradient vector for all neighborhood pixels. (17) = i wk U 101 with U. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY boundaries to be smooth. 6) K-means Clustering: The K-means clustering [45] is an iterative method used to partition an image into K clusters.


Morphological gray-scale reconstruction methods are used to eliminate noise while preserving the nuclei shape [24]. For example. cancerous and noncancerous regions are refined using morphological operations. i. Magee et al. Alternatively. A common equation is Specimen value − Background value . (21) While the emission pattern captures the tissue-dependent staining. 103 Different color models can be used. Sertel et al. [19] proposed a framework for classification of intraductal breast lesions as benign or malignant using the cellular component. Several nuclei detection and segmentation methods [25]. Adaptive filters [92]. a blank (empty) image is captured and used to correct images pixel by pixel [73]. For nuclei level feature computation. the excitation pattern captures the illumination. They proposed a rank function which maps the intensity ranges across all pixels. Morphology cannot distinguish the nuclei areas and artifacts having a nuclear-like shape but different intensity values. While thresholding and filtering reduce noise according to pixel intensities. [25] performed automated separation of cancer from noncancerous regions (stroma. Cataldo et al. Kothari et al. background.. E(i). [61]. Most detection and segmentation methods [9]. Ruifrok et al. [54]. then a trimmed average of the brightest pixels can be used to estimate the excitation pattern  (i) = EAVE J  1 Ij (i) J −K +1 (22) j =K where K is set to an integer closest to J(1 − g) + 1. including histogram or quantile normalization in which the distributions of the three color channels are normalized separately. as characterized by a structuring element. and the emission pattern. Alternatively. Gamma correction [17]. [84] explains how virtually every set of three colors can be separated by color deconvolution and reconstructed for each stain separately. and histogram equalization [52] have been used to increase the contrast between foreground (nuclei) and background regions. Macenko et al. of the image is formed from stained tissue (nonzero background). [59]. [24]. even in the presence of the object [75]. and Luv are sometimes used [11]. Anisotropic diffusion is used to smooth nuclei information without degrading nuclei edges [52]. thus. applying the threshold function on a group of pixels instead of an individual pixel eliminates a noisy region. [17]. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY the illumination pattern from a series of images.: METHODS FOR NUCLEI DETECTION. Later. Reinhard et al. [18]. The intraductal breast lesions contain four . they fail at eliminating large artifacts [91]. the preprocessing step selects the ROI by excluding regions with little content and noise [91]. Then. It requires prior knowledge of color vectors (RGB) of each specific stain. [70]. The pixels that lie outside threshold values are often determined using intensity histogram are considered to be noisy. [83]. [25]. [67]. [86] are using color deconvolution-based separation of stains in histopathological images. color deconvolution is effective in separation of stains [84]. Dalle et al. [86]. [19]. a method further refined by Niethammer et al. Thresholding is popular for ROI detection. blood vessels) using unsupervised clustering. [86]. Clustering is another method that is commonly used for ROI detection. Similarly. White Reference value − Background value (20) A downside of this method is that a blank image must be acquired for each lens magnification whenever the microscope illumination settings are altered.e. Alternatively. morphological operations can also be used for noise reduction. Noise and artifacts are eliminated using morphological operations like closings and openings [59]. [26]. While such techniques are successful to eliminate small spots of noise. Ij (i) denotes an ordered set of pixels. [81] used histogram-based normalization in histopathological images. [49]. [86]–[89]. Other more perceptual color models such as HSV. [85] and Magee et al. It applies a separate linear normalization for each pixel where class membership is defined by a pixel being colored by a particular chemical stain or being uncolored. although the RGB model is not a perceptually uniform color model. 2) Color Normalization: Many color normalization techniques have been proposed [78]–[81]. [72]. noise reduction is succeeded by ROI detection to determine the nuclei region [70]. [83] extended Reinhard’s normalization approach to multiple pixel classes by using a probabilistic (GMM) color segmentation method. In white shading correction. Thresholding (prior or subsequent to applying the morphological operations) removes such artifacts. Dundar et al. [10]. [82] proposed a method for matching the color distribution of an image to that of the reference image by use of a linear transform in a perceptual color model (Lab color space). for tissue level feature computation. a very common phenomenon in histopathological images. Assuming that a certain percentage. [52] introduced the nuclei and cytological components as ROI for grading of follicular lymphoma (FL). 3) Noise Reduction and Image Smoothing: Thresholding is used for noise reduction that usually follows filtering and background correction in order to minimize random noise and artifacts [22]. Another possible way is to estimate background by exploiting the images of the specimen directly. morphology reduces noise based on the shape characteristics of the input image. [51]. g. 4) ROI Detection: In some frameworks. [55]. thresholding is performed in RGB color model for elimination of RBCs and background. Gaussian filtering is also used to smooth nuclei regions [18].IRSHAD et al. Red blood cells (RBCs) and background regions show uniform patterns as compared to other nuclei in FL tissue. noise reduction and ROI detection are performed simultaneously. [76]. [90]. SEGMENTATION. [17] selected neoplasm ROI for nuclei pleomorphism in breast cancer images by using Otsu thresholding along with morphological operations. [64] use the RGB color model. [77] introduced a method to correct nonuniform illumination variation by modeling the observed image I(i) as product of the excitation pattern. [80] proposed the automatic derivation of these color vectors. An alternative normalization method is based upon the intrinsic properties of the image which are revealed through Gaussian smoothing [74]. From a set of J images. [27]. M (i) as Transmittance = I(i) = E(i) × M (i). Lab. [50]. [70]. In order to deal with stains colocalization. Can et al.

sin(θ(x. 2) Use the Euclidean distance map DN (i) to constrain the maximum scale values when combining the LoG filtering results across scales to compute a single response surface RN (i) as RN (i) = arg max {LoGnorm (i. y. [59] also employed RST for seed detection. .and intrareader variations. The accuracy of segmentation methods depends critically on the reliability of the seed points. σ) is a 2-D shifted Gaussian kernel defined as   1 (x − μx )2 + (y − μy )2 exp − . is a prerequisite for most nuclei segmentation methods. ξm ax ]. the false positive rate (FPR) is reduced by four times by using this technique [86]. 2014 components: cellular. C. y. μy and standard deviation σ located at the center (x. The shifted Gaussian kernel is specifically designed by amplifying the voting at the center of the targeted object and resulted in low occurrence of false seeds in overlapping regions. y. [64] proposed a novel and fast algorithm for seed detection by utilizing single-path voting with the shifted Gaussian kernel. μx . usually one per nucleus and close to its center. Nuclei Detection The identification of initial markers or seed points. Parameters of the GMM model are estimated using EM [46]. several other approaches have been proposed to detect the seed points. ξ) × IN (i)} (23) ξ ∈[ξ m i n . μx . and Δ is the aperture angle of the cone. y) is computed using the negative gradient direction −(cos(θ(x. The voting direction α(x. Using textural information. 7. The remaining pixels are considered cellular region and are used in lesion classification. σ) (u . The Hough transform detects seed points for circular-shaped nuclei but requires heavy computation [49]. Δ). rm ax . compute the response of the scale-normalized LoG filter (LoGnorm (i. y) is the magnitude of gradient image and N (x. . Δ) =  I(x. μy .104 IEEE REVIEWS IN BIOMEDICAL ENGINEERING. The voting image V (x. y. manual counting of nuclei is tedious and subject to considerable inter. Although this approach is inspired by the results of the generalized symmetry transform. Veta et al. where θ is the angle of the gradient direction with respect to x-axis. The Centroid transform also detects seeds but limitations make it useful only for binarized images. extra cellular. rm in . including computational efficiency and ability to exploit shape and size information. and illumina. Δ) is generated using the shifted Gaussian kernel with its means μx . The radial symmetry transform (RST) is also used for seed detection. The H&E-stained image data are modeled into four components using GMM. Later. rather than considering the contribution of a local neighborhood to a central pixel. The main steps of this methodology are as follows. ξ) = ξ 2 LoG(i. [58] proposed a distance-constrained multiscale LoG filtering method to identify the center of nuclei by exploiting shape and size cues available in the Euclidean distance map of the binarized image. Δ) with its vertex at (x. 3) Identify the local maxima of RN (i) and impose a minimum region size to filter out irrelevant minima. Those classified as the cellular component are further clustered by dynamic thresholding to eliminate blue–purple pixels with relatively less luminance. The Euclidean distance map is commonly used for seed detection and Laplacian of Gaussian (LoG) is a generic blob detection method. being unable to exploit additional cues. Initial works in this field rely upon the peaks of the Euclidean distance map [17]. Alkofahi et al. rm in . the seed points are determined by executing mean shift on the sum of voting images. μy . y) is used to define the voting area A(x. . y)). it determines the symmetrical contribution of each pixel around it. y. Counting nuclei by type is highly important for grading purpose. a cone shape (rm in . The H-maxima transform detects local maxima as seed points [26]. rm ax . y) of the voting area A and oriented in the voting direction α using single path approach as V (x. μx . [88] proposed the fuzzy c-means (FCM) clustering method along with the ultraerosion operation in the Lab color model for detection of MN in IHC images of meningioma. MN count provides clues to estimate the proliferation and the aggressiveness of the tumor [62]. μy . They reported detection accuracy N (x. . They also reported intrapathologist error of 21. rm ax . hypo-cellular stroma (HypoCS) and hyper-cellular stroma (HyperCS). ξ)) at multiple scales ξ = [ξm in . [53]–[55]. σ) = . 2 × DN (i)}}.2%. This approach employs magnitude and phase spectrum in the Gabor frequency domain to segment HypoCS and HyperCS regions. The resulting mixture distribution is used to classify pixels into four categories. 1) Initially. Using multiscale LoG filter with a Euclidean distance map offers important advantages. Fuchs and Buhmann [95] reported 42% disagreement between five pathologists on classification of nuclei as normal or atypical. VOL. where rm in is a minimum radius. First. rm ax . Khan et al. Recently. For MN detection in breast cancer histopathology images. Qi et al. Anari et al. The main disadvantage of this methodology is its sensitivity to even minor peaks in the distance map that results in over segmentation and detection of tiny regions as nuclei. min{ξm ax . y. Loy and Zelinsky [93] proposed fast gradient-based interest operator for detection of seed points having high radial symmetry. and IN (i) is the nuclear channel image extracted by separating the foreground pixel from background pixel using automatic binarization. y)). respectively. However. This shows the high potential added value of automatic counting tools.ξ M A X ] where ξM AX = max{ξm in . They have compared their results with iterative voting method in [94]. regions with hues of red. 2πσ 2 2σ 2 (25) where μx = x + cos2 θ (rm ax + rm in ) and μy = y − sin2 θ (rm ax + rm in ). [70] proposed a novel and unsupervised approach to segment breast cancer histopathology images into two regions. rm in . This methodology improves the accuracy of seed locations. y) N (x.v )∈A (24) where I(x. being highly sensitive to texture and often resulting in overseeding. rm ax is a maximum radius.

In order to estimate unknown parameter (θ). AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY nearly equal as manual annotation. . C2 . . such that Ui=1 vk i = 1.: METHODS FOR NUCLEI DETECTION. The loglikelihood function of parameter vector θ is defined as (26) (θ) = k =1 i=1 with m > 1 (m ∈ R). . The membership function vk i is vk i = U j =1 with the cluster center Ck = 1 I (i)−C k I (i)−C j . . SEGMENTATION. U is the total number of pixels in I. C = {C1 . The FCM clustering method is based on the minimization of the following objective function: Jm (V. the EM method is employed for the maximum likelihood estimation. C) = c  U  vkmi I(i) − Ck 2 σ. and V = [vk i ] is a c × U matrix in which vk i is the kth membership value of ith pixel.IRSHAD et al. Cc } are the cluster centers.

In this case. For instance. σ)]} (33) i=1 Recently. Nuclei Segmentation Nuclei features such as size. Roullier et al. some feature computation method requires the exact boundary points of nuclei to compute the nuclei morphology. mitotic regions are initially segmented by using the following discrete label regularization function:   λ 0 2 f −f min R(f ) + (29) 2 f ∈H(V ) where the first term R(f ) is the regularizer defined as the discrete Dirichlet form of the function f ∈ H(V ) : Rw (f ) = 1 2 12 and H(V ) is the u ∈V [ v ∼u w(u. pixel intensity of mitotic and nonmitotic region is modeled by a Gamma–Gaussian mixture model as f (Ii . θ) where f (I(i). σ) logf (I(i). In this framework. 2 are indicator variables showing the component membership of each pixel I(i) in the mixture model (31). The second term is a fitting term. μ. resulting in a large number of nuclei other than lymphocytes being detected. Cires¸an et al. and other morphological appearance are important indicators for grading and prognosis of cancer. Grading of lymphocytic infiltration based on detection of large number of LN in IHC HER2+ breast cancer histopathology was reported by Basavanhally et al. The use of EM for GMM was recently proposed by Khan et al. v)(f (v) − f (u)) ] 2 Hilbert space of real valued functions defined on the vertices V of a graph. ψ. N (I(i). U m i=1 vk i 105 (31) where ρ1 and ρ2 represent the mixing proportions (prior) of the intensities belonging to mitotic and nonmitotic regions. This method reported 51% F-score during ICPR 2012 Contest [96]. k = 1. λ ≥ 0 is a fidelity parameter called the Lagrange multiplier which specifies the tradeoff between the two competing terms. ξ)]} i=1 {wi2 log[N (Ii . m 2−1 c (θ) = 2 U   U (32) wik logρk + + U  U  {wi1 log[Γ(Ii . respectively. More details on these definitions can be found in [62]. High detection sensitivity has been reported for this framework. LN are automatically detected by a region growing method which uses contrast measures to find optimal boundary. shape. v) (30) where f (t) is function at the iteration step t. v)f (t) (v) ⎩ f (t+1) (u) = . This framework has been evaluated on 41 HER2+ WSI and reported 90. θ) = ρ1 Γ(I(i). [86] for the detection of MN in breast cancer histopathological images. ξ) represents the Gamma density function for mitotic regions. ∀u ∈ V λ + v ∼u w(u. The choice of the nuclei segmentation method is correlated with the feature computation method. The EM method finds the maximum likelihood estimation of the marginal likelihood by iteratively applying expectation and maximization steps iteratively as (27) (28) U  i=1 i=1 k =1 m i=1 vk i · I(i) . σ) represents the Gaussian density function for nonmitotic regions. classification and grading of cancer is highly dependent on the quality of segmentation of nuclei. This approach consists in unsupervised clustering at low resolution followed by refinements at a higher resolution. [97] used deep max-pooling convolutional neural networks (CNNs) to detect MN and achieved highest F-score (78%) during ICPR 2012 contest [96]. The Gauss–Jacobi method is used to approximate the solution of minimization in (29) by the following iterative algorithm: ⎧ (0) 0 ⎨ f (u) = f (u) λf 0 (u) + v ∼u w(u.41% detection accuracy as compared to 94. ξ) + ρ2 N (I(i). Γ(I(i). it is parameterized by shape (ψ) and scale (ξ) parameters. At multiresolution level. In order to reduce the number of FP. In this framework. size and luminance information based maximum a posteriori (MAP) estimation is applied to temporarily labeled candidates as either LN or CN. Their approach proved to be very efficient and to have a much lower number of false positives (FPs) as compared to the other contestants. Markov random field (MRF) theory with spatial proximity is used in order to finalize the labels. The trained CNN is then used to compute a map of probabilities of mitosis over the whole image. μ. θ) is the mixture density function in (31). it is parameterized by μ and θˆ = argmax (θ) (34) θ where wik . ψ. [62] proposed a graph-based multiresolution framework for MN detection in breast cancer IHC images. high magnification images are required to utilize the exact . A training dataset consisting of patch images centered on ground truth mitosis is used to train a CNN. This discrete regularization is adapted for labeling the mitotic regions at higher resolution. texture.59% manual detection accuracy. The authors reported more than 70% TPR and 80% TNR. μ. Later. [18]. ψ. D. Consequently.

∀j = 1. Finally. [98]. Initially. Cosatto et al. a simple methodology to segment nuclei [9]. Ta et al. a biquadratic filtering is used to produce a gradient image from IEEE REVIEWS IN BIOMEDICAL ENGINEERING. . v) = min ρ∈P G (u . In addition.). . K and j . [54]. Dalle et al. Furthermore. GiPS reports overall 7. u). transformation into polar coordinate system is performed for every patch with the center of mass of the nucleus as the origin. The reported results are suboptimal for ring-shaped nuclei having clear homogeneous regions. Huang and Lai [24] proposed watershed and ACM-based framework for nuclei segmentation in hepatocellular carcinoma biopsy images. A pseudometric δ : V × V → R is defined as δ(u. [89]. The simplest way to detect and segment nuclei in histopathological images is based on thresholding and morphological operations. . Then. watershed. nuclei are detected using thresholding and morphological operations. and fitness parameters is used to extract nuclei boundaries. ACM is applied to generate smooth and accurate contours for nuclei. However. A large number of publications on nuclei segmentation in histopathology use state-of-the-art image segmentation methods based on thresholding. [49] described an automated method for accurately and robustly measuring the size of neoplastic nuclei and providing an objective basis for pleomorphism grading. the resulting segmentation is strongly dependent upon the initial seed points. a dual morphological gray-scale reconstruction method is employed to remove noise and accentuate the shapes of nuclei. 2. The edge map and distance transform are used for seed detection [26]. the energy δ : V → R induced by the metric δ for all the seeds of S can be expressed as δS (u) = min δ(si . using multidimensional kernels. making it harder to find a reliable threshold level. 2. ACMs can combine both shape characteristics (smoothness and shape model) with image features (image gradient and intensity distribution). and Gcuts. preferably uniform background. Finally. Initially. it can be extended to color images or stacks of images. Finally. Then. The difference between nuclei and background regions may be diffuse. a novel image partition (graph reduction) algorithm is proposed for segmentation of nuclei in serous cytological and breast cancer histopathological images. it requires the prior detection of seed points. ACMs. the watershed transform does not include any prior knowledge to improve its robustness. ui+1 ) (f (ui+1 ) − f (ui )) i=1 (35) where w(ui . . This methodology actually suffers from its simplicity by including little object knowledge. The specificity of this framework is to use graphs as a discrete modeling of images at different levels (pixels or regions) and different component relationships (grid graph. K. noisy background. and damaged/irregular nuclei. Based on Voronoi diagrams. texture. an ACM with shape. s i ∈S ∀u ∈ V. 7. the Hough transform is used to pick up radially symmetric shapes. However. morphology.84% accuracy error. The authors claimed 90% TPR. separately or in combination. . a difference of Gaussian (DoG) filter is used to detect nuclei.106 details of nuclei. a marker-controlled watershed transform is performed to find the edges of nuclei. . [99]. proposed gradient in polar space (GiPS). [15]. Then. clustering. it lacks robustness on size and shape variations. Given a set of K seeds S = (si ⊆ V). [10]. The main advantage of watershed is that there is no tuning to do before using it. . This framework achieves poor segmentation in case of low contrast. This methodology is not meant to segment clustered or overlapping nuclei. a novel nuclei segmentation method [17]. v) is a set of paths connecting two vertices.v ) m −1   w(ui . The main parameters to tune are the threshold level and the size and shape of the structuring elements. where i = 1. VOL. This methodology reports higher performance on well-defined. 2014 which nuclei boundaries are delineated. proximity graph. First. Other feature computation methods require their course location to compute topology features. which are very frequent in histopathological images. region growing. It can be defined. (36) The influence zone z (also called Voronoi cell) of a given seed si ∈ S is the set of vertices which are closer to si than to any other seeds with respect to the metric δ. as well as on texture variations. [53] proposed a method based on graph-based regularization. . etc. [54]. Several authors have been using the watershed transform for nuclei segmentation [26]. ui+1 ) is a weight function between two pixels and PG (u. Even though this methodology is usually defined only on gray-scale images.

I  (j)) E(I  (i)) = − ln P(I(i)) + i j ∈N (i)   I(i) − I(j) × exp − 2σI2 (38) where P(I(i)|k). if I  (i) . The pixel labeling I  (i) is done by minimizing the following energy function:   η(I  (i). is the set of influence zones Z(S. as z(si ) = {u ∈ V : δ(si . This method reported 95. The authors compared this method with kmeans clustering and Bayesian classification methods in [100].47% accuracy. the foreground is extracted using Gcut-based binarization. N (i) is a spatial neighborhood of pixel i. δ) = {Z(si ). [58] proposed another Gcuts-based method for segmentation of breast CN. Initially. u) ≤ δ(sj .= i. (37) Then. for a given set of seeds S and a metric δ. 1 is a Poisson distribution.73% segmentation accuracy as compared to k-means clustering and Bayesian classification methods which reported 93. Kofahi et al. respectively. and  1. the energy partition of graph.67% and 96. k = 0. ∀si ∈ S}. u)} .

In (38). I  (j)) = 0. After binarization. the first term is a data term that represents the cost of assigning a label to a pixel and the second term is a pixel continuity term that penalizes different labels for neighboring pixels when |I(i) − I(j)| < σI  . These detected seed points are used to perform initial segmentation which . otherwise. nuclear seed points are detected by combining multiscale LoG filtering constrained by a distance map-based adaptive scale selection (23).= I  (j) (39) η(I  (i).

they are intolerant to chromatin variations. The magnetostatic active contour [103] model is used as a force F guiding contour toward boundary. v and u are L(v) and L(u). By minimizing a morphological-based cost function. smooth boundaries. [51] proposed a nuclei separation methodology in which concave vertex graph and Ncut algorithm are used.IRSHAD et al. It is computed as follows:  θ(cw ) = π − arccos (cw − cw −1 ) · (cw +1 − cw ) |cw − cw −1 ||cw +1 − cw |  (41) where cw is a point on the contour. The authors reported 86% accuracy on 25 histopathological images containing 7400 nuclei. Region growing from seed points produces initial segmented nuclei. Initially. The degree of concavity/convexity is proportional to the angle θ(cw ) between contour points. [66] proposed a multireference Gcuts framework for solving the problem of technical and biological variations by incorporating geodesic constraints. The authors claimed 80.: METHODS FOR NUCLEI DETECTION. The authors report 95% accuracy with computational cost of one second per field of view image. Later.3% accuracy. The main problem with most ACMs is their sensitivity to initialization. these frameworks have poor segmentation accuracy for CN especially when CN are clustered and overlapping. In addition. the outer boundary is delineated via robust estimation and color active model. statistical and shape models are used to separate overlapping and clustered nuclei. L(u)) is the prior energy. L(u)) (40) (v . respectively.u )∈E where Eg f and Elf are the global and local data fitness terms applying the fitness cost for assigning L(v) to v. However. resulting in the removal of poorly contrasted objects. distance transform and shape-based cluster separation methodologies are applied keeping only the separation lines. The authors reported 85% TPR and 75% PPV on TCGA dataset [101] of 440 WSI. and clustered nuclei by using machine learning algorithms together with classical segmentation methods.59% accuracy in case of overlapping nuclei and 95. [26] addressed the problem of clustered nuclei and proposed a methodology that combined the intensity and gradient information along with shape parameters for improved segmentation. [60]. Then. homogeneous chromatin distribution. overlapping. Then. seeded watershed segmentation is applied on the gradient magnitude image to create the region borders. which went through deep valleys in the distance map. In case of highly clustered nuclei with weak borders between nuclei. and individual existence. and differences of the staining. E. combined global and local thresholding are used to select foreground regions. The outputs of these two detectors are merged using an ACM to refine the border of the detected nuclei. As compared with nuclei segmentation methods. Then. First. Cloppet and Boucher [99] presented a scheme for segmentation of overlapping nuclei in immunofluorescence images by providing a specific set of markers to the watershed algorithm. these methods are more tolerant to variations in shape of nuclei. At last. morphological filtering is applied to detect seed points. partial occlusion. Wahlby et al. They defined markers as split between overlapping structures and resulted in 77. SEGMENTATION. [54]. MC. and fibroblast nuclei in IHC breast cancer images [69]. Vink et al. Yang et al. and EN having regular shape. The framework often causes oversegmentation when chromatin is highly textured and the shape of nuclei is extremely elongated. high concavity points are detected on the contours and used in the construction of an edge-path graph. a similar approach is used for segmentation of clustered and overlapping nuclei in tissue micro array (TMA) and WSI colorectal cancers. These nuclei segmentation frameworks have reported good segmentation accuracy on LN. Initially. and a concave vertex graph is constructed from automatically detected concave points on boundaries (41) and inner edges. Furthermore. To solve this initialization problem. EM-based ACM initialization allows the model to focus on relevant objects of interest. Based on contours enclosing multiple objects. the optimal path in graph is recursively calculated to separate the touching nuclei. For nuclei segmentation in glioblastoma histopathology images. During labeling. The first detector focuses on the inner structure of nuclei and second detector covers the line structure at the border of nuclei. clustered nuclei are separated using watershed and ellipse approximation. a scheme based on high concavity points and size heuristic is used to resolve overlapping nuclei. . LN. which are very common in CN. Nuclei Separation A second generation of nuclei segmentation frameworks tackles the challenges of heterogeneity. Morphological filtering is used for finding nuclei seeds. In final step. [25]. The authors reported 90% accuracy for overlapping nuclei. and Esm o othness (L(v). denoting the cost when the labels of adjacent vertices. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY is refined later using a second Gcuts-based method with combination of alpha expansion and graph coloring to reduce computational complexity. the authors report that one detector cannot cover the whole range of nuclei as diversity in appearance is too large to be covered by a single detector. [57] proposed EM-driven Geodesic ACM with overlap resolution for segmentation of LN in breast cancer histopathology and reported 86% TPR and 64% PPV. Chang et al. In [102]. introduced a deterministic approach using machine learning technique to segment EN. 107 The watershed transform is employed to address the problem of overlapping nuclei by defining a group of basins in the image domain. the result of the initial segmentation is refined with gradient magnitude along the boundary separating neighboring objects. a unique label L(i) is assigned to each vertex v ∈ V and the image cutout is performed by minimizing the energy E=  (Eg f L(v) + Elf L(v)) v ∈V +  Esm o othness (L(v). [19]. They formulate two detectors (pixel-based and line-based) using modified AdaBoost. Fatakdawala et al. undersegmentation may occur. where ridges in-between basins are borders that isolate nuclei from each other [9].83% overall accuracy.

each pixel is classified into nuclei or background regions by utilizing colortexture in the most discriminant color model. The watershed transform is used for model initialization.j . Hχ 1 ∨χ 2 = (Hψ 1 + Hψ 2 − Hψ i Hψ 2 ). [64] proposed a two-step method for the segmentation of overlapping nuclei in hematoxylin-stained breast TMA specimens that require very little prior knowledge. First. a modified GAC with the Chan–Vese energy model is used to detect the nuclei region [104]. for the recognition of single nuclei in nuclei cluster. The authors evaluated this framework on overlapping nuclei in prostate and breast cancer images and reported 86% TPR and 91% OR on breast images and 87% TPR and 90% OR on prostate images. Last. Ψ. region homogeneity. (c) Level set [105]. φ2 ). B}. ω > 0 are constants that balance contributions of the shape and boundary. 4. H(·) is the Heaviside function. respectively. the inner edges are extracted by applying the watershed transform on a hybrid distance transform image. Qi et al. Hχ 1 ∧χ 2 = Hψ 1 Hψ 2 . the boundaries of touching clumps are smoothed out by Fourier shape descriptor and then concave point detection is carried out. [60] integrated a framework consisting of a novel supervised nuclei segmentation and touching nuclei splitting method. βr . [28]. [63]. Moreover. the following level set representation of the contours is used: E = αN K   k =1 +β Λk K  1  k =1 +λ |I − μk | di + αB 2 K   k =1 |I − μb |2 di ΛB g(|I(k (z))|)|k (z)|dz 0 K K   Λk ∩ Λj (43) k =1 j =1. Another adaptive AC scheme that combines shape. and δ(φk ) is the contour measure on {φ = 0}. proposed an automatic separation method for clustered nuclei in breast cancer histopathology [61]. For initial segmentation of nuclei. and Θj = |I − Ij |2 + μ|Ij |2 and j ∈ {F. (a) Original (b) GIPS [17]. boundary. First. The differentiation between clustered and separated nuclei is computed using the distance between the radial symmetry center and the geometrical center of the connected component. Fourth. IF and IB are foreground and background regions. ψ2 ). points are computed by executing mean shift on the sum of the voting images (24). Kong et al. VOL. high concavity points along touching nuclei regions are detected (41). the concave vertex graph using high concavity points and inner edges is constructed. Results of segmentation and separation using different methods on same area of an image.108 IEEE REVIEWS IN BIOMEDICAL ENGINEERING. βs . Second. IB ) = βs K =2  k =1   (φk (I) − ψ(I))2 |φk |δ(φk )dI   Shape + boundary energy   (ΘF Hχ 1 ∨χ 2 )dI + (ΘB − Hχ 1 ∨χ 2 )dI βr +     Region energy  K =2   ω Hχ 1 ∧χ 2 dI + (φk − ψk )2 dI  + k =1     Mutual occlusion energy  (42) where Φ = (φ1 . Second. seed Fig. IF .55% splitting ER. the optimal separating curve is selected by computing the shortest path in the graph. Third. The authors evaluated this framework on FL images and achieved average 77% TPR and 5. 2014 Mouelhi et al. 7. Ψ = (ψ1 . δ(·) is the Dirac delta function. (d) MPP [65]. which combines the geometric distance and color gradient information. region and mutual occlusion term. and mutual occlusion terms in a multilevel set formulation was proposed by Ali et al. The segmentation of K overlapping nuclei with respect to shape prior ψ is solved by minimizing the following level set φ function: E(Φ. For splitting of clustered nuclei.

K is the number of nuclei. β > 0 are constants that balance contributions of each term. The repulsion term separates the touching nuclei to create smooth and complete contour of each nuclei. is able to segment overlapping nuclei as several individual objects. [65] proposed a method based on marked point processes (MPPs). The resulting modality has a strong object-background contrast and smoothing out the irregularities within the nuclei and background. Fig. To overcome ACMs initialization sensitivity. Later. . [67] proposed a method based on the creation of a new image modality consisting in a gray-scale map where the value of each pixel indicates its probability to belong to a nucleus. segmentation is performed using an ACM with a nuclei shape prior [65] to solve the problem of overlapping nuclei. There is no need to initialize the process with seed points giving the location of the nuclei to be segmented. 4 shows a comparison of nuclei segmentation results using MPP. scale information. The authors claimed 78% TPR and 90% PPV in case of touching nuclei. This methodology.= k where αN . . A shape prior term is used for handling overlapping nuclei. This probability map is calculated from texture. Recently. 5(a) shows the result of ACM segmentation on . K) is the nuclei contours that evolve toward boundaries. Kulikova et al. where ν controls the slope of the output curve and ζ controls the window size. αB . μk and μb are mean intensities of nuclei and background regions. . k (k = 1. Fig. ΛB is the background which represents the regions outside all the nuclei. . Λk is the region inside each contour k . and g is a sigmoid function x −ν g(x) = (1 + e( ζ ) ). and levelset [105]. Veillard et al. and simple pixel color intensities. GiPS [17]. The last term in (43) is the repulsion term used to represent the repulsion energy between each touching nuclei and λ is a regulation parameter. a type of high-order ACM.

and limited ability to segment multiple overlapping objects. Fifth. One limitation of RST is the prior knowledge of scale. A prior knowledge for the overlapping nuclei is incorporated to obtain separation line without jaggedness. After the final iteration. In general. which cannot be generalized. [56] proposed adaptive likelihood-based nuclei segmentation for FL centroblasts. To address this problem. This framework avoids a segmentation bias with respect to shape features. they achieved 93. Sertel et al. maxima are used as marker of a nuclei segmentation algorithm such as watershed. rm ax and the angular range Δ. Furthermore. and 3. The main idea is to have nuclei as compact and as well separated as possible. while a large number of steps increases computational cost. (d) ACM on Hematoxylin stained image. as well as to reconstruct occluded contours in overlapping region. all initially generated contours are ranked using gradient fit.80%. rm in . Few number of steps resulted in the fragmentation of the center of mass. y) − 2 2  (44) where the radial range is parameterized by rm in . nonoverlapping segmentation is performed with ranked labeling in a 2-D map. [59] proposed a method similar to [24] that met the objective of nuclei segmentation in H&E-stained breast cancer biopsy images by applying the fast RST [93] to produce markers for the watershed segmentation. nuclear components are clustered using GMM with EM. Veta et al. y) + θ(x. conditional erosion. Then. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY Fig.and GMM-based unsupervised Bayesian classification scheme was used for segmentation of overlapping nuclei in IHC images [55]. 5. [68] proposed a novel contourbased minimum model for nuclei segmentation using minimal a prior information. rm ax . cluster parameters are chosen to maximize ϕε . segmentation is improved using contour optimization. θ(x. y) is the angle between the positive xaxis and the voting direction. Δ Δ ≤ φ ≤ θ(x. First. (b) ACM on probability map image. These methods are mainly dependent on the availability of expert annotations.: METHODS FOR NUCLEI DETECTION. In . 5. The voting area can be regulated by selecting the number of steps in the evolution of the kernel shape. Initially. local maxima locations associated with individual nuclei are determined. (c) Hematoxylin stained image. Fourth. RST is an iterative algorithm attributing votes to pixels inside a region [93]. Second. to allow robust nuclei segmentation. The separation of overlapping nuclei is formulated as cluster analysis problem. Schmitt and Hasse [106] separated the clustered nuclei using RST based on the idea that the center of mass in a nucleus is considered as a basic perceptual event that supports separation of clustered nuclei.43% with respect to classical watershed. These parameters are updated using votes from the previous iterations. They initialized iterative voting along the gradient direction where. all possible closed contours are computed regardless of shape and size. Third. The authors managed to achieved 86% TPR and 91% PPV on a dataset of 7931 nuclei. Nuclei Features and Classification Features computed from segmented nuclei are usually a prerequisite to nuclei classification that generate higher level information regarding the state of the disease. This minimum model-based segmentation framework consists of six internal processing steps. (a) Probability map image. which capture the deviations in the nuclei structures. and adaptive H-minima transform schemes in terms of separation accuracy. y. at each iteration.70%. Segmentation results using ACM methods on probability and hematoxylin-stained image [67]. the voting direction and shape of the kernel are refined iteratively. SEGMENTATION. Δ) = (x + r cos φ. segmented regions are classified as nuclei or background using stained related information. Each boundary point contributes to votes for a region defined by oriented cone-shape kernels as  A(x. They initially obtained regions of clustered nuclei by clustering and levelset segmentation. multiscale extension of the RST seems to be reasonable. probability map image and hematoxylin-stained image. sensitivity to initialization. The classifiers use nuclei features. the spatial voting matrix is computed along the gradient direction. Alternatively. They achieved improvements of up to 6. which is viewed as a mixture of Gaussian. Overall. model-based approaches segment nuclei using a prior shape information. EM. Finally. Using fast RST. Thus. To overcome this limitation. The cluster validity index consists of a compactness measure ϕ (the smaller value means more compact) and a separation measure ε between the clusters. Last. F. these models may not be generalizable and have limited application due to the manual training step. produced after color deconvolution [84]. Recently. y + r sin φ) |rm in ≤ r ≤ rm ax . The novelty of these approaches corresponds to the use of machine learning and statistical methods to eliminate malformed nuclear outlines and thus. cluster nuclei are separated using concavity point detection (41).IRSHAD et al. cluster validation is performed to evaluate the optimal number of nuclei. A similar method [106] is used in [50] to decompose regions of clustered nuclei in H&E-stained prostate cancer biopsy images. Wienert et al.48% segmentation accuracy for overlapping nuclei on specimens of cervical nuclei and breast invasive ductal carcinomas. They also proposed a way 109 to deal with holes and sub holes in the region by processing boundaries iteratively. a parametric EM algorithm is employed to learn the distribution of topographic surface (GMM). This approach primarily involves applying the distance transform to generate topographic surface. On the basis of extracted regional maxima. which may introduce a bias favoring the segmentation of nuclei with certain characteristics. to learn how to classify nuclei into different classes.

Using selected features after excluding highly correlated features. [19]. the accuracy rate of classification promoted from 92. the combination of several texture measures instead of using just one might improve the overall accuracy. These datasets should be medically validated. In addition. A summary of nuclei features is listed in Table III. Moreover. yet the RL and FD performed better in high and low shape frequency. but found relatively little such work for histopathology imagery. They also studied the texture measures response to additive texture distortion noise while varying nuclei shape densities. [72] computed intensity. the GMRF was nearly uniform. Intensity and morphological features are extensively used for nuclei classification as epithelial and CN in [17]. and run-length (RL) features in selective color channels from different color models with a decision tree and SVM classifiers for mitosis detection in MITOS dataset of breast cancer histopathology images and ranked second with 72% F-score in ICPR 2012 contest [96]. morphological. the computed features.54% with an SVM-based decision graph classifier. which comprise both local and global characteristics so that benignancy and different degrees of malignancy can be distinguished effectively. The combination of GMRF and RL improved the overall accuracy up to 92. They computed four different texture features. the MITOS mitosis . while the RL and FD performed better in the high frequencies. An SVM-based decision graph classifier with feature subset selection on each decision node of classifier is used in comparison with k-nearest neighbor and simple SVM. They extracted a total of 14 features (intensity. One of the aforementioned challenges is the lack of unified benchmarks. To address the problem of heterogeneity in CN. In histopathology. a significant number of articles have been published in the field of histopathology. [89]. various evaluation methods. For numerical comparison of the studies. these features can be categorized into the following four categories: 1) cytological. [72]. and morphological features and used these features with SVM for the classification of segmented candidate regions into mitotic and nonmitotic regions. focusing on nuclei segmentation and classification in different image modalities. which should be covered in future research. Still. [57]. Recently. VOL. CO.50% with none of the classified meningioma subtypes achieving below 90%. These open research areas have unique challenges. This method reported 66% F-score during ICPR 2012 contest [96]. and texture features) from segmented nuclei in biopsy images. We found a compilation of features for cytopathology imagery [107]. morphological.88% to 94. Daubechies wavelets. morphology. Malon et al. According to Al-Kadi [10]. They studied the variation of texture measure as the number of nuclei increased. [67] used intensity and textural features with support vector machine (SVM) classifier for the creation of a new image modality to segment CN. Their results of feature selection demonstrated that Zernike moment. In some frameworks. we know of only a few benchmark datasets: UCSB Bio-Segmentation [111]. [69] constructed a large set of features and modified AdaBoost to create two detectors that solved the problem of variations in nuclei segmentation. there are two types of information available in the image: 1) the intensity values of pixels. and 2) their spatial interdependency [29]. Different texture measures tend to extract different features each capturing alternative characteristics of the examined structure. Irshad et al. V. the second detector is constructed using Haar like features. and annotated by different pathologists to accommodate subjective variations in annotation. Definitions for all listed features can be found in [29]. Nguyen et al. Vink et al. respectively. it is not straightforward to evaluate and numerically compare different studies solely based on their reported results since they use different datasets. it is definitely necessary to build benchmark datasets. the other two are statistically based: CO and RL features. [98]. An exhaustive set of features including morphological and texture features are explored to determine the optimal features for nuclei classification [109]. By observing the cancer detection procedure adopted by pathologists. and 4) texture features. DISCUSSION AND FUTURE DIRECTIONS Since the last decade. Huang and Lai [24] extracted intensity and cooccurrence (CO) features. Prominent nucleoli (cytological feature) inside nuclei region is used to classify nuclei as cancerous or not. 7. like intensity and texture features. and multiple performance metrics.110 IEEE REVIEWS IN BIOMEDICAL ENGINEERING. Recently. two of them are model-based: Gaussian Markov random field (GMRF) and fractal dimension (FD). Veillard et al. Studies cited in this review have been performed using their own private datasets. The first detector is formulated with intensity features. and textural features having 78% TPR on a dataset including six WSI for training and 11 for testing. and [49]. The GMRF was the least affected. Bayesian classifier was trained for meningioma subtype classification. texture. [110] used intensity. are explicitly used for segmentation of nuclei with K-means clustering [56]. In addition to the morphological features computed from cytological regions. To the best of our knowledge. comprise samples coming from a large number of patients. [27] developed a novel idea for cancer detection in prostate using cytological (nuclear) and textural features. intensity. prostate cancer is detected using cytological. and [108]. 3) morphological. 2) intensity. 2014 TABLE III SUMMARY OF NUCLEI FEATURES USED IN HISTOPATHOLOGY order to extract features. Such an effort would make possible the numerical comparison of the results obtained by different studies and the identification of distinguishing features. there are some open research areas where little study has been done. Similarly. and Gabor wavelets are the most important features for nuclei classification in microscopy images.

Brat. 2011. Mag. [58]. 74–77. USA. Biomed. Street. A. it has not been addressed well in the literature. Thompson. Zhou and S. 43.” IEEE Signal Process. [9] M. H. Boston. 30–Sep. “Nuclear feature extraction for breast tumor diagnosis. “Robust cell nuclei segmentation using statistical modeling. [4] X. and J. “Morphological feature extraction for the classification of digital images of cancerous tissues. S. [2] M. pp. cellular. Conf.2%. E. and B. R. marker-controlled watershed needs true nuclei markers. 2010.” in Proc. [11] J. Aug. H.. [3] R. Dec. Biol. Mitotic count is an important parameter in breast cancer grading as it gives an evaluation of the aggressiveness of the tumor. Int. and 4) magnification. [60]. Eng. in Rubin’s Pathology: Clinicopathologic Foundations of Medicine. no. Wong. and tracking. Histol. Although the seemingly large figures are to be interpreted into the specific context of the study.. Rajpoot. GMM and EM [55]. E. [113] reported a moderate agreement between three pathologists for identifying MN on H&E-stained breast cancer slides. and A. pp. accurate segmentation of touching or overlapping nuclei is still an open research area. Gorstein. pp. Conf. “Informatics challenges of high-throughput microscopy.” Sci. 59. distance transform [11].” Bioimaging. Strayer. [55]. SVM [67]. Adjei. 302. 1996. [19]. A variety of schemes taking into account concavity point detection [28]. Al-Kadi. and J. A conclusion of this study is that that self-assessment is not a reliable validation method. “A better lens on disease: Computerized pathology slides may help doctors make faster and more accurate diagnoses. The MITOS benchmark has been set up as an international contest of mitosis detection in the framework of conference ICPR 2012. REFERENCES [1] W. Mouroutis. vol.–3 Sep. Tasks include segmentation. 1905. 10. Int. the problem has not yet been completely solved. pp. [59].. 2006. no. 6. [10] O. Symp. Detection of mitosis is a very challenging task. Cytol. vol. A. since mitosis are small objects with a large variety of shape configurations. [8] T. Med. Saltz. P.. The basic philosophy of the machine learning approach is that human provides examples of the desired segmentation. 14. F. the overall best results on mitosis detection presented during the recent MITOS and AMIDA contests achieved an F-score of 78. The two main avenues to be explored in terms of supervised machine-learning algorithms are the use of more domain specific features and limitation of overfitting issues. N. and ellipse-fitting techniques are unable to accommodate the shape of most nuclei. “Image analysis for neuroblastoma classification: Segmentation of cell nuclei. [24]. Can. Ma and A. To the best of our knowledge.” in Proc. USA. 111–137. 2009. no. region growing tends to rely on shape and size of nuclei. and J. J. vol. Technol. L. 2. pp.” CA: A Cancer J. 2004. AMIDA benchmark reedited in 2013 the same type of mitosis detection challenge as MITOS did in 2012. Eds. vol. [67]. Bibbo. Oct. as well as a recent similar initiative AMIDA [112]. vol. The data cover subcellular. They employ only local intensity information without any prior knowledge about the object to be segmented and produce inaccurate nuclei boundaries. While different methods have been developed with various levels of success in the literature for the problem of overlapping and clustered nuclei. Rubin. [25].” Pattern Recog. Kong. vol. [57]. N. 2. 6605–6608. which would not be considered accurate detection under medical terms. pp. classification. 6. however. NY. S. Eng. These methods have good results for nuclei that are slightly touching or overlapping each other. Most of these methods also require prior knowledge. N. vol. Rubin and D. there are still progresses to be made to reach clinically acceptable results. 2010. “Histopathological image analysis: A review. W. E. “Novel agents on the horizon for cancer therapy. Soc.and intrapathologist disagreements is also to be taken into account. Med. Bartels. It is also important to address the issue of robustness to varying clinical/technical conditions including: 1) different scanners used for image acquisition. 79–91. Wolberg. [7] J. Strayer. Gurcan. 43. L. pp. Thiran and B. 3. A. The UCSB Bio-Segmentation Benchmark dataset consists of 2-D/3-D images and time-lapse sequences that can be used for evaluating the performance of novel state-of-the-art computer vision methods. Saltz.21% for MITOS [71] and 61. 6. D. it shows that validation by medical expert is not a straightforward issue. adaptive ACM with shape and curvature information [63]–[65]. [26]. A. M. Apr. the computation of curvature is highly dependent on concavity point detection algorithm. and new image modality [67] able to deal with heterogeneity. L. [68]. pp. For instance. Macq. and tissue level.. SEGMENTATION. May 2006. and AdaBoost [69] are used for nuclei segmentation.. 23. M. and O.. Pan. S. IEEE 28th Annu. Yener. 2009. T. 1998. PA. Shimada. These methods suffer from dependencies inducing instability. Philadelphia.. 3. and leaves the optimization and parameter tuning tasks to the learning algorithm. D. 31 Aug. [51]. only few supervised machinelearning techniques like Bayesian [18].: METHODS FOR NUCLEI DETECTION. Schwarting. IEEE 33rd Annu. S. W. 3) different lightening conditions. Cooper. pp. USA. Boucheron. marker-controlled watershed [9]. pp. In spite of the availability of few methods like clustering. vol.. Quant.” Anal. no. . The issue of inter. and watershed can locate the nuclei region but problems arise when they try to segment the touching and overlapping nuclei.IRSHAD et al. Dealing with overlapping and clustered nuclei is still a major challenge in the field of nuclei segmentation. Biol. 63–72. 4844–4847. Eng.1% for AMIDA [112]. 861–870.: Sci. [54]. 1992. 147–171.. Eng. MA. Segmentation methods like thresholding. [12] W. For instance. “Towards building computerized image analysis framework for nucleus discrimination in microscopy images of diffuse glioma. 459–473. 1011–1020. “Bayesian belief networks in quantitative histopathology.. Bharath. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY detection [71] benchmark. N. Imag. Weber. [61]. Clinic. GMM and EM. USA: Lippincott Williams & Wilkins. and graphs [51]. Kurc. Most of these benchmarks highlighted the fact that despite the promising results. mainly because of the lack of available data. They also reported intrapathologist error of 21.” in Proc. [6] P. Biomed. New York. 1993.” IEEE Trans. region growing. Int. May. 2) different staining characteristics. Madabhushi. Mangasarian. “Texture measures combination for improved meningioma classification of histopathological images. Soc. M. T. vol. Amer. CA. and D. 4th ed. pp. Fuchs and Buhmann [95] reported 42% disagreement between five pathologists on nuclei classification as normal or atypical. [5] M. Roberts. 2043–2053. no. Electron. H. [61] have been 111 investigated to separate overlapping and clustered/touching nuclei. San Jose. A similar study by Malon et al. 1–4 Feb. but they are not suitable for specimens containing larger numbers of nuclei with extensive overlapping and touching. Gurcan. The MITOS benchmark has been set up to provide a database of mitosis freely available to the research community.” IEEE Rev.

Bilgin. vol.” Comput. 2007. 15th Int. “Blobworld: Image segmentation using expectation-maximization and its application to image querying.. 20. Catalyurek. Veltri. “Combining intensity. Med. San Diego. Racoceanu. 2002. Eng. J. P. Jain. IEEE Workshop Appl. Sci. pp. C. E. J.” in Proc. D. Tampa. X. 1977–1984. L. 382–399.” Microsc. D. “Nuclear pleomorphism. 321–331. Statist. “Automated cancer diagnosis based on histopathological images: A systematic survey. pp. Agner. EPC1993-110. N. H. Leow. vol. Fox. 1113–1125. Basavanhally. MacQueen. Biomed. 1. pp. Putti. “Maximum likelihood from incomplete data via the EM algorithm. 3. K. 107–112. 2002. no. Bhanot. [54] C. Orlando. Jung and C. Elston and I. W. D. “Effective segmentation and classification for HCC biopsy images.. [21] N.. 57.” Int. Image Process.” Cancer Res. Dalle. 3. and B.” Histopathology.” Histopathology. pp. Comput. 3. 6 pp. pp. 2001. vol. no.” Sequ. 2008. Intell. 1. and M. no.. [14] C. Xiong. Interv. 7.. 642–653. Gurcan. Madabhushi. Conf. H. K. Eng. Biol.” in Proc. 57. 2009. 1976. Roerdink and A. Gurcan. pp. Reden. 215. [44] J. Miller. and A. Demir and B. 41. F. Biomed. 1997.-T. 5th Berkeley Symp. pp. [47] C. L. FL. C. Comput. 731–737. Bhanot. ON. Summer School Morphological Image Signal Process.” Bioinformatics. 661–669. [28] S. Med. 6. vol. pp. Comput. Epstein. Tomaszewski. Monaco. Levine. Sertel. V. 6. Sanya. D.” Computer. 4. Graph. 365–372. Int.. E. 192. Image Analysis and Mathematical Morphology. pp. [45] J.: Methods Appl. Eng. 42. P. pp. Feldman. R. “Computeraided prognosis of neuroblastoma: Detection of mitosis and karyorrhexis cells in digitized histological images. Huang. “Scoring nuclear pleomorphism in breast cancer. Sertel. V. Jul. “Point set morphological filtering and semantic spatial configuration modeling: Applications to microscopic image and bio-structure analysis. “Computerized image-based detection and grading of lymphocytic infiltration in HER2+ breast cancer histopathology. W.” IEEE Trans. Basavanhally. Ganesan. 7–27. Jain. pp. N. Bhanot. 39. no. K. T. pp. pp. Racoceanu.. K. “High expression of lymphocyte-associated genes in node-negative HER2+ breast cancers correlates with lower recurrence rates. pp. pp. “A novel shape prior based segmentation of touching or overlapping ellipse-like nuclei. “Image analysis and morphometry in the diagnosis of breast cancer. S.. “Mathematical morphology: Basic principles. M. 2010. Eng. VOL. F. pp. 43. 2. [42] J. Acquaviva. 2825–2832. Xu. pp.” J. B. no. pp. Pathol. “Adaptive thresholding for the digital desk. vol. 2. 2004. the value of histological grade in breast cancer experience from a large study with long-term follow-up. Wahlby. no. [48] J. Biomed. 2000. 1979. Carson. Hainan. [22] S. pp.. Pattern Recog. [26] C.” Pattern Recog. [38] P. 4. 2011. pp. Conf. USA: Academic. and M. Aug. vol. [25] S. E. Soc.. vol. Badve. vol. [20] C. Racoceanu. Comput. [30] L. Katsinis. G. Weber. Malik.-H. 1–4. 109– 118. “Snakes: Active contour models. Zucker. [52] O. “Prostate cancer detection: Fusion of cytological and textural features. 19th Int. DeLisi. Oct. Yener. 1991. 57. D. pp.” Fundamenta Informaticae. [31] H. IEEE 31st Annu. “Grading nuclear pleomorphism on histological micrographs. 349–356. vol. no. and T. Canada. H. D. M. Treatment. Roy. [29] C. 579–591. no. vol. USA. [39] J. USA. Singh. N. Gurcan. C.” IEEE Trans..” in Proc. Poland.” Pattern Recog.” in Proc. W. He. 67. Laxminarayan. Mach. Eng. Vis. Informat. Cretu. [53] V. 2008. and D. Foran. 109–116. [41] S. “Image segmentation with implicit color standardization using spatially constrained expectation maximization: Detection of nuclei. 2014 [34] C.. J. Sep. 2010. Terzopoulos. 351–362. Oct. 2011. 8. “Automatic image analysis of histopathology specimens using concave vertex graph. pp. Tuzel. Huang. M. Lai. Informat. 281–297. Imag. 145–151.. and E. N. pp. [49] E. Garcia. “The watershed transform: Definitions. Minnesota. S. vol. 10. 14th Int. pp. Series B (Methodological). 2010. Mar. Bharucha. 2011. Rubin. 53. A. Jung. FL. NY. no.. S. [24] P. 57. 8. P. vol. Jul. M.” Breast Cancer Res. Vis.” J. 24. P. Malik. “Nuclear pleomorphism scoring by selective cell nuclei detection. L´ezoray. Conf. Ficarra.” in Proc. Rosen. 2008. 14 pp. [32] J. no. Dec. Microsc. 1433–1436. Suri. 1. 8 pp. Int. 187–228. 10. pp. “Computer-aided detection of centroblasts for follicular lymphoma grading using adaptive likelihood-based cell segmentation. Shi and J.” Pattern Recog. 1.” in Proc. Fatakdawala. Genome Anal. Going. “Adaptive energy selective active contour with shape priors for nuclear segmentation and gleason grading of prostate cancer. 1977. no. C. Meijster. Elmoataz. vol. Hipp. Statist. M. Dec. vol. Eng. N. Lozanski. Biomed. “Computerized classification of intraductal breast lesions using histopathological images. A. 57. USA. 1–4. Conf. 2000. 2600–2604. 151–153. E. Gil. 63. “The cell graphs of cancer. Comput. M. P. Biomed.” Dept. 100. China. Interv. CA. “Large-scale computations on histology images reveal gradedifferentiating parameters for breast cancer. S. 1995. [37] N. U. N. Barnard. S. Ali.. no. 2012. 6914. 5. Liu. 259–265. vol. Tampa. Hafiane. Biomed..-Assist. He and Q.. Ta. H. pp. NY. Ellis. G. Conf. Wang. vol.112 [13] J. and L. Huang and Y. 2012. 3. Probabil. TR-05-09. 58. Monaco. S. [19] M. no. J. “Region growing: Childhood and adolescence. S. 1. Stierer.-H. M. 18– 22 Sep. Image Comput. [43] M. H. Martel. “Automated segmentation of tissue images for computerized IHC analysis. 2009. Glenn McCluggage. “An automatic threshold selection method based on discriminate and least squares criteria.” in Proc.” IEEE Trans. Hamilton.. Nov. I. A.. N. Interv. Pattern Anal. Dempster. and S. no. 7. G. Nguyen. [33] G.. and G. “Clustering initiated multiphase active contours and robust separation of nuclei groups for tissue segmentation. Putti. Methods Progr. J. 41. O. vol. [16] A. Conf.. A. “Unsupervised segmentation of overlapped nuclei using Bayesian classification. W. Sintorn. [40] H. Zakopane. 2011. no. Belongie. [18] A. vol.” J. pp. no. Graf. 2009. Eng. and A. Feldman. Tech. [27] K.” Denshi Tsushin Gakkai Ronbunshi. Conf. and A. Wu. Gultekin. and J. Oh. H. Puerto Rico. Haber. Chae. Kass. 2004. vol. Eng. . Palaniappan.-K. [36] B. CA.-Assist. Image Comput. 1993. W. Laird. no. Tech. vol. Petushi. 2010. and A. J. G. 833–841. Meyer. H.” J. Dec. Otsu. [56] O. Keenan. Diamond... J.” in Proc. 38–40. J. San Juan. Mar. algorithms and parallelization strategies. “Expectation maximization-driven geodesic active contour with overlap resolution (EMaGACOR): Application to lymphocyte segmentation on breast cancer histopathology. 8–28. no. “Normalized cuts and image segmentation. no. W. 271–287. 1967. Dunne and J. FL. 1550–1563. U. 1–38. Liao. 1983. S. Math. A. Borgefors.” in Proc. Bartels.. 7. pp. Balis. Scanfeld.” in Proc. Christudass. no. 2002. and P. Heijmans. Comput. W..” in Xerox. Imag.-R. P. Image Comput. Biomed. A. IEEE REVIEWS IN BIOMEDICAL ENGINEERING. A. A. Lom´enie and D. Pathol. Inf. Meer. F. Madabhushi. pp. [35] M. J. pp. 2002. Thoma. vol. H. 20. Racoceanu. Greenspan.” IEEE Trans. and D. J. Vis. USA.. “Knowledge-guided semantic indexing of breast cancer histopathology images. Pattern Recog.. and M. [15] S. Antani. [17] J. Lucas. 2010. and R. Soc. vol. [46] A. B. vol. Pathol. Shimada. S. U. and E. vol. Biomed. and S. pp. G. and B. vol. Technol. Gunduz. “Time-efficient sparse analysis of histopathological whole slide images. USA. C. Raykar. “Is H&E morphology coming to an end?. vol. J.” IEEE Trans. Med. no. USA. pp.” J. 2. Clin. pp. O. M. Lom´enie. Tomaszewski.” IEEE Trans. “Shape recovery algorithms using level sets in 2-D/3-D medical imagery: A state-of-the-art review. and D. [50] A.. O. B. Toronto. Res. Thompson. 19th Int. Berkeley. and D. Rep. [57] H. Dec. Sep. Li. Sch¨upp. Leow. H. 2010. Macii. Smith.” IEEE Trans.. Roux. P. Med. 1–15. S. [23] X. “Graph-based tools for microscopic cellular image segmentation. Zeng. 1026–1038.. SPIE. S. H. 39. Bunyak. New York. Rensselaer Polytechnic Inst. 10 669–10 676.. 1676– 1689. IEEE Comput. edge and shape information for 2D and 3D segmentation of cell nuclei in tissue sections. Wellner. Y. C.” IEEE Trans. Shana’ah. 6. Veillard. pp. G. and S. 1.. and V. Witkin. Pattern Recog. Graph. [51] L. 2.” in Proc. USA. Biomed. pp. Yang. no.” BMC Med. D. 2005. 2. 2008. S. 1988. 1–18. N. Madabhushi. and J. Ganesan. 45. a strong prognostic factor in axillary node-negative small invasive breast cancer. “Computer assisted diagnosis in histopathology. 67–76. Sertel. Soc. W.. Alexe. Serra. 2010. Kim. S. Harris. vol. Cataldo. vol. Med. and G. D. 2008. Troy. 2613–2616. vol. “Pathological prognostic factors in breast cancer. 59. pp. Long. pp. L. vol. pp. Tutac. and K. 2000. Erlandsson. R. 12.. 2010. Kim. Dundar. 35. S. Yener. Cosatto. R. “An automated machine vision system for the histological grading of cervical intraepithelial neoplasia (CIN). Dalgin.” Comput. [55] C. 14–24. 2006..” in Proc. vol. Comput. vol.-Assist. Mesirov. 2894– 2911. “Some methods for classification and analysis of multivariate observations. 22. C. and A.. pp. “Segmenting clustered nuclei using h-minima transform-based marker extraction and contour parameterization. Tozeren. Sabata. pp. Madabhushi. Tamayo. Ganesan. Bengtsson.

[73] G.. and M. and N. Kulikova. M. Imag. Endosc. Marron. Quirke. P. M.: Nano Macro. 2010. J. 4.” in Proc. 2011. Gerdes. Macenko. 657–660. L. FL. Biomed. Sun Valley. Woods. Reinhard. Jul. 7. W. BarcellosHoff.” in Proc. L. 6–7 Jul. Veillard.. Roullier. and M. C. Imag. 291–299. U.” in Proc. Inform. X. Leong. Soc. Intell. Irshad. Pathol. J. Anderson.. Piccinini. Naour. [72] C. Aug. NJ. “Computational pathology: Challenges and promises for tissue analysis. 475–486. MITOS. Niethammer. and A. 618–621. 2011. Amer. Danielsen. Paris./Int. 2001. [100] O. A. “Iterative voting for inference of structural saliency and characterization of subcellular events. Anal. A. S. “Fast radial symmetry for detecting points of interest.” J. Eng. 2008. 42. vol. [71] (2012). D. Pathol. no. Technol. 59.. SEGMENTATION. J. pp. Conf. Symp.” IEEE Trans. L´ezoray. Montalto. Jul. A. Int. N. H. D. 57. pp.K. K. 1–5. P. Capron. Loss. Conf. 2012. Microsc. Pathol. Tunisia. Isfahan. Belkacem-Boussaid. USA. no. 2010. J. 7. 9–13. 2011. Young. [94] B. Image Process. 603–615. 3 pp. “Transillumination method apparatus for the diagnosis of breast tumors and other breast lesions by normalization of an electronic image of the breast. and D. “Improved automatic detection and segmentation of cell nuclei in histopathology images. Chang. Med. May 2013. pp. 2003. Biol. San Diego. H. 2008. Albregtsen. D. [88] V.” IEEE Trans. Symp. Acad. A. Mahzouni. Malon and E. no. C./Oct. Tampa. Crellin. L.” J. V. B. T. O. no. pp. Grenier.” in Proc. 149–152. and N. no. IEEE Int. Biomed.. Venice. pp. Stenzinger. Heim. and G.. H. V. 25. May 2013.. O. A. R. M. Med. “Automatic batch-invariant color segmentation of histological cancer images. and F. B.. Marron. Boucher. and P.” IEEE Trans. D. vol. Bello. 16. Haan. Japan. Schmidhuber. Klossa. Treanor. C. Symp. Tissue Image Anal. Lucarelli. Guan. C. Image Process. Roysam. 1999. M. Pattern Recog. “Colour normalisation in digital histopathology images. P. Chicago.” J. Veta. pp. Cloppet and A. Image Comput. USA. Imaging Graph. Borland. 4. Saeger. Nagoya. J. Lassoued. Elmoataz. Cytol. [65] M. Woosley. 2010. Telepathol. Zelinsky. [83] D. Racoceanu. Loy and A. 1997. Lomenie. M.. Biomed. [78] L. and J. V. 2. 959– 973. “A method for normalizing histology slides for quantitative analysis. Comput.” in Proc. Med. M. 2012. 5th Int. Control Appl. Leeuwen. 100–111.. Thomas. Hammamet. Giusti. Virt.K. W. M. Treanor. M. Irshad. Anari. [95] T. “Quantification of histochemical staining by color deconvolution. and J. 714–720. Image Process. 124– 135. and K.. M. USA: Pearson. Baheerathan.. Mar.. [85] M.. pp. pp. W. “A hybrid magnitude-phase approach to unsupervised segmentation of tumor areas in breast cancer histology images. Digital Image Processing. Ciresan..” in Proc. 8. Yang.” in Proc. 1. 8–14. 6th Iran. . W. [70] A. Kong.” IEEE Comput. 31. 2012. and B. T. “Multi-modal imaging of histological tissue sections. 2011. Brady. Marty.. Beil. Gonzalez and R. 2002. Mar. Sood. 2001. Cytol. vol. Xing.: Nano Macro. 2007. no.” in Proc. vol. shapebased active contour for multiple object overlap resolution in histological imagery.. Med.. Conf. pp. [69] J. 35. E.. and D. [86] A. 2011.cnrs. D. [63] S. “New texture features based on the complexity curve. 3. 2013. France. U. texture. 3. Madabhushi. 605–618. Boston. H. vol. “Mitosis detection in breast cancer histological images an ICPR 2012 contest. A. A. Bartels. Vis. pp. “Automatic segmentation of clustered breast cancer cells using watershed and concave vertex graph. vol. Montironi. B. USA. Mach. D. Microsc. Interv. P. Irshad. 411–418. Intell. 30. scale and shape information. IEEE Int. 2. Sayadi.. 515–530. Khan. Comput. N. 7 pp. E. Mach. L. pp. Foran. Phan. Al-Kofahi. pp. Xiaodong. NY. “A Gamma–Gaussian mixture model for detection on mitotic cells in breast histology images. H.. Yang. Mouelhi. A. 19th Int. vol. D. pp. Chaudry. “Appearance normalization of histology slides. Genestie. Soc. 35. Parvin. Iran. 2012. Q. [76] F. IEEE 35th Annu. Image Understand.” in Proc.. Mar. Imag. 288–291. vol. Comput. M. and F. pp. pp. F. Dietel. 9. Inform. M. Simmons. “Automated location of dysplastic fields in colorectal histology using image texture analysis. vol. London. Rep. 11–15 Nov. and F. no. 56.. pp. Clin. 8. Biomed. vol. Kulikova. Bevilacqua. Piccinini.: Nano Macro. “Partitioning histopathological images: An integrated framework for supervised colortexture segmentation and cell splitting. Sloan. vol. Mar. 2011. “Mitosis detection in breast cancer histology images with deep neural networks. Comput.: Nano Macro. Viergever. J.” in Proc. E. pp. 1–6. 2013. Adhikhmin. Huisman.” J. and B. 841–852.” in Proc. Borland. vol. Opt.” in Proc. 2009. Denkert. no. no. V. 7. “Cooperation of color pixel classification schemes and color watershed: A study for microscopic images. M. Graph. [89] H. 1–6. [74] F. Idaho. Deurzen. 2012. Eng.” IEEE Trans. F. Japan. Tsukuba. C.. pp. USA.. Racoceanu.” BioTechniques. “Multi-resolution graph based analysis of histopathological whole slide images: Application to mitotic cell extraction and visualization. M. Quirke. Pathol. 113 [79] D. Med. Meer. Cardot.. “Context aware colour classification in digital microscopy. M. pp. vol. and J. Congr. Jun. pp. Sep. Can. Osaka. Shires. Commun. and D. R. and L. Racoceanu. Italy. Available: http://ipal.” IEEE Trans. 2012. 1–5. 58–66. M. Barcelona. Mohee. Roux. pp. Hamilton. Ruifrok and D. Coventry. Gherardi. Microsc. “Automated mitosis detection in histopathology using morphological and multi-channel statistics features. Gurcan. Moffitt. Rajpoot. Pattern Anal. [96] L. Apr. McIntosh. Biomed. Magee. 4.. M. P. Ta. Rydberg. vol. Inf. 2005. Woosley. Histol. [77] A. 11th Eur. pp. “Cell nuclei extraction from breast cancer histopathology images using color. 9th IEEE Int.. “Nuclear segmentation in H&E sections via multi-reference graph cut (MRGC). [67] A. “Color transfer between images. Amirfattahi. 619–621. A. Pluim. 1661–1677. Imag. [62] V. 68–75. [97] D. [90] S. [60] H. Bioinform. K. New York. E. no. 615–623. McGee. vol. “Illumination field estimation through background detection in optical microscopy. Imag. Gherardi. USA.” in Proc. 6–9. Wienert. H.IRSHAD et al. 249. and J. no. “Detection and segmentation of cell nuclei in virtual microscopy images: A minimum-model approach. 32. “Multi-image based method to correct vignetting effect in light microscopy images. Sep.” Pattern Recog. “Robust segmentation of overlapping cells in histopathology specimens using parallel seed detection and repulsive level set. Yang. 8th IEEE Int. Gooch.. “Marker-controlled watershed segmentation of nuclei in H & E stained breast cancer biopsy images. Gurcan. J. Stokes. M. USA: Springer. Sep. 182. pp. 5.” Comput.” in Machine Learning in Medical Imaging. 1107– 1110.” in Proc. van Diest. Biomed. pp. P. Pathol. D. Chang. no. and P. vol.” IEEE Trans. 6–22. Roux. 2012. 9. [64] X. 23. [Online]. 1–4. pp. Gambardella.. M. A. Microsc. “Computer-aided detection of proliferative cells and mitosis index in immunohistichemically images of meningioma. boundary-. Schmitt. IEEE 5th Int. 11. El-Daly. Med. H. H. 21th Int.” Comput. 2003. [84] A. Kothari. Inform. Orth. Thomas. 2007. “Unsupervised segmentation based on robust estimation and color active contour models. Vink. Q. 2010. “Efficient nucleus detector in histopathology images. Biol. vol. E. 22–26 Sep. pp. and N. 2011. Ali and A. 21.. Appl. pp. Veillard. and M. “Multi-channels statistical and morphological features based mitosis detection in breast cancer histopathology.” in Proc. Ginty.. CA. 1448–1460. and M. 8–11 Dec. [75] A. 20–24 Sep. [98] H.. G. [87] L. 2011. 7. H. Cytol. Med. Imag. pp. Hassberger. Spain. H..” J..” presented at the Proc. Congr. J. Kulikova. Khan. MA. W. Inform. “Classification of mitotic figures with convolutional neural networks and seeded blob features. C. M. pp. pp. Eng. [59] M. vol. Han. Lezoray and H. E.-Assist. Shirley. Histopathol.. J. Roux. pp. Cosatto. Chomphuwiset. Japan. [99] F.. F. vol.” Sci. IPAL UMI CNRS Lab Std. 248. Magee. Symp. “Nuclei extraction from histopathological images using a marked point process approach. Funt.” IEEE Trans.” J. 7 pp. pp. F. Pattern Recog. 2008. Imag. H. Biomed. 754–765. pp. Thompson. ICPR 2012 Contest. Symp. 6091–6094. M. Smith. T. “Blank-field correction for achieving a uniform white background in bright field digital photomicrographs. 30 Aug–3 Sep. J. 3–7 Jul.. N. vol. and P. 614– 617. IEEE Symp. AND CLASSIFICATION IN DIGITAL HISTOPATHOLOGY [58] Y. Conf.: Nano Macro. N. D. Upper Saddle River. T.” U. 4. K. A. Graph. E. vol. A. Buhmann.” J. [61] A. M.” Anal. no. 3. vol. [66] H. [82] E. “Segmentation of overlapping/aggregating nuclei cells in biological images. D. E. Biomed. and N. Qi. “An integrated region-. Rajpoot. 2013.: METHODS FOR NUCLEI DETECTION. [68] S. 6. M. no. Cline.” IEEE Trans. Bevilacqua. Conf. Lee. L. Imag. D. “Correction of uneven illumination (vignetting) in digital microscopy images. pp. [92] R. [93] G.. and A. 10–15. and D. May 2013. SPIE Med. 2012. 2013.. Johnston. D. 783–789. Foran. no. 34–41. and H. P. J. P. [81] S. Fnaiech. E. 4. C. [80] M.” J. 4. 4. pp. Jan. Klauschen. El-Daly. Patent 5 079 698. 1992.. [91] P. Hufnagl. Imag. vol. and R. Int. Racoceanu. Wang. Pathol. and D. no. IL. Fuchs and J. Apr. Quant. Parvin.S. Niethammer.

Lahore. 2.” IEEE Trans. IEEE REVIEWS IN BIOMEDICAL ENGINEERING.” Anal. 12–18. and B.. E. 14. Obara. Eng.. Liu.” J. in particular. Image Process. Med. degree at the University Joseph Fourier. no. Irshad. Dept. classification. Chan and L. and 2006. dissertation. French National Center for Scientific Research. P.S.” in Proc.. 3–7 Jul. [109] S. Brachtel. Huang. Singapore. Saltz. “Automated mitosis detection using texture. France. 5445– 5448. [108] L. pp. [111] E. 863– Comput. 7. webpages). Hasse. degree in computer science from the National University of Computer and Emerging Sciences. French National Center for Scientific Research Lab. vol.. AMIDA. Japan. 2008. [110] H. medical image analysis. USA. Yagi. vol. J. He is currently working toward the Ph. S.uu.” IEEE Trans. [Online]. Available: http://amida13. 97– [113] C. Ludovic Roux received the Ph.” BMC Bioinform. G. IEEE 35th Annu. pp. Shu. Kvilekval. F. 30–Sep.S. Grenoble.) at the National University of Singapore. Singapore. in 2005 and the M. Feb. E. Fedorov. 30. Singapore. R. and Dr. vol. . 1905–1923. degree in computer science from the University Paul Sabatier. L. He is with the joint French–Singaporean Research Lab Image and Pervasive Access Lab UMI. L. Roux. D. A. in 1997 He is currently with the University Joseph Fourier.. 25. 2009. R. 266–277. CA. pp. Paris. 4.D. and the M. Rajapakse. Available: https://tcga-data. in 2009. He is the Director of the French–Singaporean-Joint Lab Image and Pervasive Access Lab UMI. IEEE 33rd Annu. and F. Cellul. Eng. France. 2013. 10. “Segmenting overalapping cell nuclei in digital histopathology images. M. Xie and M. in 1992. Joo-Hwee. 1997. Schmitt and M. Sharp. no. SIFT features and HMAX biologically inspired approach. in 2009. French National Center for Scientific Research. [Online]. Toulouse. and J. 1–12.comp. Mosaliganti. G. Aug. and J.nus./Aug. 632–646. Graph. and M. [112] (2013). and a member of Image and Pervasive Access Lab UMI.. MA.jsp [102] tcgaAnalyticalTools. “A biosegmentation benchmark for evaluation of bioimage analysis methods. Conf. Vese. vol. habil degrees from the University of Besanc¸on.. Med. Humayun Irshad (S’13) received the B.” Pattern Recog.. Manjunath. Rodenacker and E. Santa Barbara. 41. Bengtsson. Mirmehdi. Pathol. 6. and is a Research Fellow at the joint French–Singaporean Research Lab Image and Pervasive Access Lab UMI. He is a Research Fellow at the University Pierre and Marie Curie. H.. pp. Electr. National Cancer Institute Std. P. degree from the National University of Singapore. B. His research interests include symbolic and connectionist cognitive exploration methods for high-content biomedical images. S. Naour. Kaye. Univ.D.” in Proc. France. [107] K. 2008. Boston. pp. Gelasca.nci. He is currently with the University Pierre and Marie Curie. Biol. pp. “Features for cells and nuclei classification.D. Eng. and data fusion. “Mitotic figure recognition: Agreement among pathologists and computerized detector. Pathol. L. Comput. [103] X. pp. A. D. Leone.D.Eng. 1–36. vol. Boucheron. France. VOL. Kurata. His research interests include machine learning and the integration of prior-knowledge into kernel machines. Mar. in 2013. MICCAI 2013 Challenge Std. Singapore. California. and computer vision. Jul. [104] T. and a Professor (adj.) degree from Ecole Polytechnique. “A feature set for cytometry on digitized microscopic images. Capron. Malon. vol. “Reconstruction of cellular biological structures from optical microscopy data. C. Soc. Meyer. Pakistan. Kuroda. Saito.Sc. He is involved in the ANR-A*STAR program and French Ministry of Industry founded grants (http://www.” Ph. 2014 Antoine Veillard received the Diplˆome d’ing´enieur ´ (M. “Object and spatial-level quantitative analysis of multispectral histopathology images for detection and characterization of cancer. Graf. 3. image processing. (2008). His research interests include pattern recognition. no.isi. degree in computer science from the University of Central Punjab. 4. French National Center for Scientific Research. and Y. France. France. Timisoara.Sc. Mach. D. [106] O. Mundra. 35. 2012. Paris. Islamabad. “Correction to “MAC: Magnetostatic active contour model” [apr 08 632-646]. Pathol. 2011. 6601–6604. 10. He is a Senior Research Fellow at the French National Center for Scientific Research. [105] K. with a double major in computer science. Jalali. L. Biol.nih. P. Machiraju. Daniel Racoceanu (M’08) received the M.” J. Romania.114 [101] Glioblastoma TCGA Dataset.. Inform. Int. Cooper. Vis. H. 2013. His research interests include machine learning. 2008. Besanc¸on. A. 5. vol.. Fu. France.. “Radial symmetries based decomposition of cell clusters in binary and gray level images. Conf. no. Paris. Pattern Anal. Intell. Singapore. K. respectively. S. “Active contours without edges. K. Grenoble. pp.. Wu. vol. 2001.” IEEE Trans. Racoceanu. Soc.. segmentation. Cosatto.. pp. Int. machine learning. Osaka. USA. in 1993. Cellul. 2003. G.D. Singapore. and the Ph. A. pp. Qiu. degree from the Politehnica University. Ilyas. May 2008. H. Pakistan.