Kwak et al. BMC Cancer 2011, 11:62 http://www.biomedcentral.com/1471-2407/11/62

RESEARCH ARTICLE

Open Access
Multimodal microscopy for automated histologic analysis of prostate cancer
Jin Tae Kwak1,2, Stephen M Hewitt3, Saurabh Sinha1*, Rohit Bhargava2,4*

Abstract

Background: Prostate cancer is the single most prevalent cancer in US men whose gold standard of diagnosis is histologic assessment of biopsies. Manual assessment of stained tissue of all biopsies limits speed and accuracy in clinical practice and research of prostate cancer diagnosis. We sought to develop a fully-automated multimodal microscopy method to distinguish cancerous from non-cancerous tissue samples. Methods: We recorded chemical data from an unstained tissue microarray (TMA) using Fourier transform infrared (FT-IR) spectroscopic imaging. Using pattern recognition, we identified epithelial cells without user input. We fused the cell type information with the corresponding stained images commonly used in clinical practice. Extracted morphological features, optimized by two-stage feature selection method using a minimum-redundancy-maximalrelevance (mRMR) criterion and sequential floating forward selection (SFFS), were applied to classify tissue samples as cancer or non-cancer. Results: We achieved high accuracy (area under ROC curve (AUC) >0.97) in cross-validations on each of two data sets that were stained under different conditions. When the classifier was trained on one data set and tested on the other data set, an AUC value of ~0.95 was observed. In the absence of IR data, the performance of the same classification system dropped for both data sets and between data sets. Conclusions: We were able to achieve very effective fusion of the information from two different images that provide very different types of data with different characteristics. The method is entirely transparent to a user and does not involve any adjustment or decision-making based on spectral data. By combining the IR and optical data, we achieved high accurate classification.

Background

Prostate cancer
Prostate cancer (PCa) is the single most prevalent cancer in US men, accounting for one-third of non-skin cancer diagnoses every year [1]. Screening for the disease is widespread and for almost a million cases a year [2-4], a biopsy is conducted to detect or rule out cancer [3]. Manually-conducted histologic assessment of tissue upon biopsy forms the definitive diagnosis of PCa [5]. This need places a large demand on pathology services and manual examination limits speed and throughput. Histologic assessment is also critical to scientific progress as it
* Correspondence: sinhas@illinois.edu; rxb@illinois.edu 1 Department of Computer Science, University of Illinois at UrbanaChampaign, Urbana, IL 61801, USA 2 Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA Full list of author information is available at the end of the article
is often the basis for research studies. Alternative methods for histologic recognition can greatly aid in alleviating workloads, assuring quality control and reducing costs [6]. There is no straightforward way, however, to aid pathology in this task and no clinical instrument is available for routine use. Hence, high-throughput, automated and objective tools for prostate pathology - both in clinical practice and in research - are needed.
Optical microscopy and automated PCa detection
Since the tissue does not have appreciable contrast in optical brightfield microscopy (Figure 1A), tissue samples are commonly stained using hematoxylin and eosin (H&E) prior to review by a pathologist. The stain is specific in limited terms - staining protein-rich regions pink and nucleic acid-rich regions of the tissue blue (Figure 1B). A pathologist is trained to recognize, from a stained tissue sample, the morphology and local architecture of glands

2011 Kwak et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Page 2 of 16

Figure 1 Staining allows visualization of tissue features. (a) an unstained image has little contrast while (b) the application of H&E stain highlights nucleic acid-rich regions as blue and protein-rich regions at pink. (c) structure of a prostate gland. It is notable that the stain is universal in that it is not diagnostic of cell type or disease. The stain serves only to provide contrast that is subsequently used by a human to recognize cell types and diagnose disease.
as well as their structural alterations that indicate disease. The specific cell type that is used to recognize glandular structures is the epithelial sub-type. In prostatic carcinoma, which comprises more than 95% of prostate cancers [5], the cells of interest are epithelial cells [7]. Epithelial cells line 3D ducts in intact tissue and, hence, appear as cells lining empty circular regions (lumens) in images of histologic sections. Patterns of distortions of lumen appearance and spacing, as well as the arrangement of epithelial cells relative to lumens, have been characterized to indicate cancer and characterize its severity (Gleason grade) [8,9]. The greater the distortion and loss of regular structure, the worse (higher grade) the cancer. Recognizing structural distortions indicative of disease is a manual pattern recognition process that matches patterns in the tissue sample to standard patterns. Manual examination is powerful in that humans can recognize disease from a wide spectrum of normal and disease states, can overcome confounding artifacts, detect unusual cases and even recognize deficiencies in diagnoses. Manual examination, unfortunately, is timeconsuming and leads routinely to variability in grading disease [8]. Computer-aided recognition of disease samples and grade patterns [10], hence, holds the potential for more accurate, reproducible and automated diagnoses [11,12]. Unfortunately, tissue samples stain variably in populations due to biological diversity, with variations in stain composition, processing conditions and histotechnologists. The net result confounds automated image analysis and human-competitive recognition of cancer has not been automated for routine use. A robust means of automatically detecting epithelium and correlating its spatial patterns to determining cancer presence is highly desirable but yet unsolved. Several efforts have been made to develop automated systems for the diagnosis and grading of microscopic

from over 1000 patients and tens of millions of pixels using tissue microarrays (TMAs). TMAs consist of multiple tissue samples of a size that assures representative sampling and allow high throughput experimentation in an efficient manner. For this manuscript, we examined two independent data sets from prostate tissue microarrays that were subjected to chemical imaging and histologic classification as outlined above. Images of the data are shown in Figure 2. While we expected the chemical imaging approach to prove useful in histologic analysis of prostate tissue, its relationship to the existing clinical practice of using H&E stained tissue in PCa diagnosis was not clear a priori. Hence, we sought to examine whether a combination of the two techniques (i.e., optical microscopy
Figure 2 IR imaging data and its use in histologic classification. (Upper row) IR imaging data (b) is acquired for an unstained tissue section (a). The data is then classified into cell types and a classified image (c) is obtained. The colors indicate cell types in a histologic model of prostate tissue. This method is robust and applied to hundreds of tissue samples using the tissue microarray (TMA) format. (Lower row) H&E (d) and IR classified (e) images of a part of the TMAs used.

Page 4 of 16

following H&E staining, and FT-IR imaging) could provide high accuracy diagnoses that could otherwise not be achieved using H&E images alone.

Overview of this work

We develop a new fully-automated method to classify cancer versus non-cancer prostate tissue samples. The classification algorithm uses morphological features geometric properties of epithelial cells/nuclei and lumens - that are quantified based on H&E stained images as well as FT-IR images of the tissue samples. By restricting the features used to geometric measures, we sought to mimic the pattern recognition process employed by human experts, and achieve a robust classification procedure that can produce consistently high accuracy across independent data sets. We systematically evaluate the performance of the new method through cross-validation, and examine its robustness across data sets. We also summarize the specific morphological features that prove to be most informative in classification.
space (lumens) inside the tissue samples. To utilize these two features and to avoid problems due to differences in the two imaging techniques, both images are first converted into binary images. Due to the binarization, the intensity transformation is not necessary. As a spatial transformation, we use an affine transformation ( f ) [41] where a coordinate (x 1, y1 ) is transformed to the (x2, y2) coordinate after translations (tx, ty), rotation by , and scaling by factor s.

x2 t x y t 2 y cos s sin sin x1 cos y1
Accordingly, we find the optimal parameters of the affine transformation that minimizes the absolute intensity difference between two images (I reference and Itarget). In other words, image registration amounts to finding the optimal parameter values
(t * , t * , * , s *) arg min I reference f (I target ; t x , t y , , s). The x y t ,t , , s
Methods We begin with a description of the computational pipeline. As noted above, a key aspect of our approach is the use of FT-IR imaging data on a serial section that is H&E-stained to enhance the segmentation of nuclei and lumens. The first two components of the pipeline are geared to this functionality, while the next three components exploit the segmented features obtained from image data to classify the tissue sample (Figure 3).

Image Registration

downhill simplex method [42] is applied to solve the above equation. An example of this registration process is shown in Figure 4. (See [Additional file 1: Image Registration] for details.)
Identification of epithelial cells and their morphologic features
Given two images, the image registration problem can be defined as finding the optimal spatial and intensity transformation [41] of one image to the other. Here, two images are H&E stained (Ireference) and IR classified images (Itarget) which were acquired from adjacent tissue samples. The IR classified image represents the FT-IR imaging data, processed as indicated in Figure 2, to classify each pixel as a particular cell type. Although the two tissue samples were physically in the same intact tissue and are structurally similar, the two images have different properties (total image and pixel sizes, contrast mechanisms and data values). Hence, features to spatially register the images are not trivial. The H&E image provides detailed morphological information that could ordinarily be used for registration, but the IR image lacks such information. On the other hand, the IR image specifies the exact areas corresponding to each cell type, but the difficulty in precisely extracting such regions from the H&E image hinders us from using celltype information for registration. The only obvious features are macroscopic tissue sample shape and empty

While a number of factors are known to be transformed in cancerous tissues, epithelial morphology is utilized as the clinical gold standard. Hence, we focus here on cellular and nuclear morphology of epithelial nuclei and lumens. These structures are different in normal and cancerous tissues, but are not widely used in automated analysis due to a few reasons. First, as described above, simple detection of epithelium from H&E images is difficult. Second, detection of epithelial nuclei may be confounded by a stromal response that is not uniform for all grades and types of cancers. We focused first on addressing these two challenges that hinder automatically parsing morphologic features such as the size and number of epithelial nuclei and lumens, distance from nuclei to lumens, geometry of the nuclei and lumens, and others (Feature Extraction). In order to use these properties, the first step is to detect nuclei and lumens correctly and we sought to develop a robust strategy for the same.

Lumen Detection

In H&E stained images, lumens are recognized to be empty white spaces surrounded by epithelial cells. In normal tissues, lumens are larger in diameter and can have a variety of shapes. In cancerous tissues, lumens are progressively smaller with increasing grade and generally have less distorted elliptical or circular shapes.

Page 5 of 16

Figure 3 Overview of System. (a, b) FTIR spectroscopic imaging data-based cell-type classification (IR classified image), is overlaid with H&E stained image (a), leading to segmentation of nuclei and lumens in a tissue sample (b). (c,d,e) Features are extracted and selected (c), and used by the classifier (d) to predict (e) whether the sample is cancerous or benign.
Our strategy to detect lumens was to find empty areas that are located next to the areas rich in epithelium. White spots inside the tissue sample can be found from the H&E image by using a proper threshold value (>200) for the intensity of Red, Green, and Blue channels, and the pixels corresponding to epithelial cells can be mapped on the H&E image from the IR classified image through image registration. Although restricting the white areas adjacent to epithelial cells, in our observations, many artifactual lumens are still present. Additionally, the size and shape of lumens are examined to eliminate such artifacts. We note that while lumens are ideally completely surrounded by epithelial cells (called complete lumens), some tissue samples have lumens (called incomplete lumens) that violate this criterion because only a part of lumen is present in the tissue sample. To identify these incomplete lumens, we model an entire tissue sample as a circle, and the white spots between the tissue sample and the circle are the candidate incomplete lumens. As did in complete lumen

detection, the same threshold value is used to identify white areas. To identify artifacts, we use heuristic criteria based on the size, shape, presence of epithelial cells and background around the areas. In addition, the distances from the center of the tissue to the white spots are examined to identify the artifacts in crescent form which resulted from the small gaps between the tissue sample and the circle fitted to the sample. (See [Additional file 1: Lumen Detection] for details.)
Nucleus Detection - single epithelial cells
Epithelial nucleus detection by automated analysis is more difficult than lumen detection due to variability in staining and experimental conditions under which the entire set of H&E images were acquired. Differences between normal and cancerous tissues, and among different grades of cancerous tissues, also hamper facile detection. To handle such variations and make the contrast of the images consistent, we perform smoothing [43] and adaptive histogram equalization [44] prior to

Page 6 of 16

Figure 4 Image Registration. H&E stained images and IR classified images are first converted into binary images. The IR classified image is overlaid with the H&E stained image by affine transformation, with the optimal matching being found by minimizing the absolute intensity difference between two images. After registration, original annotations (color and/or cell-type information) of each image are restored.
nuclei identification. Nuclei are relatively dark and can be modeled as small elliptical areas in the stained images. This geometrical model is often confounded as multiple nuclei can be so close as to appear like one large, arbitrary-shaped nucleus. Also, small folds or edge staining around lumens can make the darker shaded regions difficult to analyze. Here, we exploit the information provided by the IR classified image to limit ourselves to epithelial cells, and use a thresholding heuristic on a color space-transformed image to identify nuclei with high accuracy. Superimposing the IR classified image on the H&E image, pixels corresponding to epithelium can be identified on the H&E image. These epithelial pixels are dominated by one of two colors: blue or pink, which arise from the nuclear and cytoplasmic component respectively. For nuclei restricted to epithelial cells in this manner, a set of general observations were made that led us to convert the stained image to a new image where each pixel has an intensity value |R + G - B|. (R, G, and B represent the intensity
of Red, Green, and Blue channels, respectively.) This transformation, followed by suitable thresholding, was able to successfully characterize the areas where nuclei are present. The threshold values are adaptively determined for Red and Green channels due to the variations in the color intensity. Finally, filling holes and gaps within nuclei by a morphological closing operation [45], the segmentation of each nucleus is accomplished by using a watershed algorithm [45] followed by elimination of false detections. The size, shape, and average intensity are considered to identify and remove artifactual nuclei. Figure 5 details the nucleus detection procedure. (See [Additional file 1: Nucleus Detection] for details.)

Feature Extraction

As mentioned above, the characteristics of nuclei and lumens change in cancerous tissues. In a normal tissue, epithelial cells are located mostly in thin layers around lumens. In cancerous tissue, these cells generally grow

Page 7 of 16

Figure 5 Nucleus Detection. Smoothing and adaptive histogram equalization are performed to alleviate variability in H&E stained image and to obtain better contrast. RG - B conversion followed by thresholding characterizes the areas where nuclei exist. Morphological closing operation is performed to fill holes and gaps within nuclei, and a watershed algorithm segments each individual nuclei. The segmented nuclei are constrained by their shape, size, and average intensity and epithelial cell classification (green pixels) provided by the overlaid IR image.
to fill lumens, resulting in a decrease in the size of lumens, with the shape of lumens becoming more elliptical or circular. The epithelial association with a lumen becomes inconsistent and epithelial foci may adjoin lumens or may also exist without an apparent lumen. Epithelial cells invading the extra-cellular matrix also result in a deviation from the well-formed lumen structure; this is well-recognized as a hallmark of cancer. Due to filling lumen space and invasion into the extracellular space, the number density of epithelial cells increases in tissue. The size of individual epithelial cells and their nuclei also tend to increase as malignancy of a tumor increases. Motivated by such recognized morphological differences between normal and cancerous tissues, we chose to use epithelial nuclei and lumens as the basis of the several quantitative features that our classification system works with. (See examples of such features in Figure 6.) It is notable that these observations are qualitative in actual clinical practice and have not been previously quantified.
Epithelial cell-related features
directly describing epithelial cells, we also quantify properties of epithelial nuclei, which are available from the segmentation described above. The quantities we measure in defining features are: (1) size of epithelial cells, (2) size of epithelial nuclei, (3) number of nuclei in the tissue sample, (4) distance from a nucleus to the closest lumen, (5) distance from a nucleus to the epithelial cell boundary, (6) number of isolated nuclei (nuclei that have no neighboring nucleus within a certain distance), (7) number of nuclei located far from lumens, and (8) entropy of spatial distribution of nuclei (Figure 6G). [Additional file 1: Epithelium-related Features] provide specifics of these measures and their calculation.

Lumen-related features

Epithelial cell information is available from IR data. However, individual epithelial cells in the tissue are not easily delineated. Therefore, in addition to features
Features describing glands have been shown to be effective in PCa classification [21,25]. Here, we try to characterize lumens and mostly focus on the differences in the shape of the lumens. The quantities we measure in defining these features are: (1) size of a lumen, (2) number of lumens, (3) lumen roundness [25], defined as L peri r where Lperi is the perimeter of the lumen, Larea 2L area is the size of the lumen (i.e., number of pixels in the

Page 8 of 16

Figure 6 Examples Features. Each panel shows one example feature, along with the distributions of the features values for cancer (red) and benign (blue) classes.
lumen), and r is the radius of a circle of size L area , (4) STD(d L cb ) lumen distortion (Figure 6A), computed as AVG(d L cb ) where d L cb is the distance from the center of a lumen to the boundary of the lumen and AVG() and STD() represent the average and standard deviation, (5) lumen minimum bounding circle ratio (Figure 6B), defined as the ratio of the size of a minimum bounding circle of a lumen to the size of the lumen, (6) lumen convex hull
ratio (Figure 6C), which is the ratio of the size of a convex hull of a lumen to the size of the lumen, (7) symmetric index of lumen boundary (Figure 6E, see [Additional file 1: Lumen-related Features]), (8) symmetric index of lumen area (Figure 6F, see [Additional file 1: Lumen-related Features]), and (9) spatial association of lumens and cytoplasm-rich regions (Figure 6D, see [Additional file 1: Lumen-related Features]). Features (3) - (8) are various ways to summarize lumen shapes,

Page 9 of 16

while feature (9) is motivated by the loss of functional polarization of epithelial cells in cancerous tissues.
Global & local tissue features

We have described above the individual measures of epithelium and lumen related quantities that form the basis of the features used by our classification system. Normally, these features have to be summary measures over the entire tissue sample or desired classification area. Hence, we employ average (AVG) or standard deviation (STD), and in some cases the sum total (TOT) of these quantities for further analysis. These features are called global features since they are calculated from the entire tissue sample. However, in some cases global features may be misleading, especially where only a part of the tissue sample is indicative of cancer. Therefore, in addition to global features, we define local features by sliding a rectangular window of a fixed size (pixels) throughout a tissue sample. For each window, AVG and/or TOT of the epithelium and lumen related quantities are computed. STD or extremal values (MIN or MAX) of the AVG and/or TOT values over all windows become local feature values (Figure 7). In all, 67 features (29 global and 38 local features) are defined capturing various aspects of tissue morphology.

Feature Selection

Feature selection is the step where the classifier examines all available features (67 in our case) with respect to the training data, and selects a subset to use on test
data. This selection is generally based on the criterion of high accuracy on training data, but also strives to ensure generalizability beyond the training data. We adopt a two-stage feature selection approach here. In the first stage, we generate a set of candidate features (Ccandidate) by using the so-called minimum-redundancy-maximalrelevance (mRMR) criterion [46] (see [Additional file 1: mRMR]). In each iteration, given a feature set chosen thus far, mRMR chooses the single additional feature that is least redundant with the chosen features, while being highly correlated with the class label. Ccandidate is a set of features that is expected to be close to the optimal feature set for a data set and a classifier under consideration. It is constructed as follows. Given a feature set F = (f1,., fM) ordered by mRMR, the area under the ROC curve (AUC) of the set of i top-ranked features is computed for varying values of i. We limit the value of i to be 30. The feature subset with the best AUC is chosen as the C candidate. In the second stage, feature selection continues with Ccandidate as the starting point, using the sequential floating forward selection (SFFS) method [47]. This method sequentially adds new features followed by conditional deletion(s) of already selected features. Starting with the C candidate , SFFS searches for a feature x Ccandidate that maximizes the AUC among all feature sets Ccandidate {x}, and adds it to Ccandidate. Then, it finds a feature x Ccandidate that maximizes the AUC among all feature sets Ccandidate {x}. If the removal of x improves the highest AUC obtained by Ccandidate, x is deleted from C candidate. As long as this removal improves upon the highest AUC obtained so far, the removal step is repeated. SFFS repeats the addition and removal steps until AUC reaches 1.0 or the number of additions and deletions exceeds 20, and the feature set with the highest AUC thus far is chosen as the optimal feature set. The classification capability of a feature set, required for feature selection, is measured by AUC, obtained by cross-validation on the training set. SFFS can be directly applied to the original feature set; however, using mRMR may help to reduce the search space and time and to build the optimal classifier by providing a good initial feature set for SFFS.

To assess the utility of the IR-based cell-type classification, we repeated the above exercises after extracting features without the guidance of the IR data; i.e., epithelial cells were predicted from the H&E images alone (see [Additional file 1: Epithelium Detection] for details). All of the features defined in Feature Extraction were used, except for Spatial association of lumens and apical regions, since the distinction between cytoplasmrich and nuclear-rich region in epithelial cells was unclear in H&E images. The results from this disadvantaged classifier are shown in Tables 1 and 2 ("feature extraction = HE only). For both types of experiments, we obtained lower average AUCs and specificity values. For instance, the AUC of cross-validation in Data2 (Table 1) dropped from 0.974 to 0.880. Similarly, the results of validation between data sets (Table 2) were substantially worse now compared to the IR-guided classification, with the AUC dropping from 0.956 to 0.918. We also observed that the average AUC dropped in the absence of IR data as using different values of parameter g for SVM (See [Additional file 1] for details). This indicates that the use of IR data, i.e., the improved epithelial identification, helps to attain better classification performance. We also note that other methods, if any, which

Page 12 of 16

could achieve high accuracy identification of epithelial cells may have the same impact with the IR data on the classification. Previously, Tabesh et al. achieved an accuracy of 96.7% via cross validation in cancer/no-cancer classification [22]. Color, morphometric, and texture features were extracted, and all images were acquired under similar conditions. We note that our classification result (Table 1), based solely on morphology, is comparable to their result; however the software developed by Tabesh et al. was not available for evaluation in our data sets. Color and texture features could provide additional information; however, their robustness to different data sets is questionable, and their interpretation is not as obvious as that of morphological features, which are used in clinical practice. Different data sets may have varied properties which may be attributable to staining variations, inconsistent image acquisition settings, and image preparation. The performance of the same method based on texture features has been seen to greatly change from one data set to another [19,22,25]. Variations in staining may affect color features. In contrast, morphological features were shown to be robust to varying image acquisition settings [25]. Nonetheless, the quality of morphological features is subject to segmentation of histologic objects. Thus, any method based on morphological features will benefit from the IR cell-type classification.

features of the tissue were quantified and employed for classification. We found that robust classification could be achieved using a few measures, which are detailed to arise from epithelial/lumen organization and provide a reasonable explanation for the accuracy of the model. The choice of combining the IR and optical data is shown to be necessary for achieving the high accuracy values observed. We anticipate that the combined use of the two microscopies - structural and chemical - will lead to an accurate, robust and automated method for determining cancer within biopsy specimens.

Additional material

Additional file 1: Supplementary material. It includes detailed description of image processing, feature extraction, feature selection, and classification method and results.
Conclusions In this manuscript, we have presented a means to eliminate epithelium recognition deficiencies in classifying H&E images for presence or absence of cancer. The method is entirely transparent to a user and does not involve any adjustment or decision-making based on spectral data. We were able to achieve very effective fusion of the information from two different modalities, namely optical and IR microscopy, that provide very different types of data with different characteristics. Several
Acknowledgements The project described was supported by Award Number R01CA138882 from the National Cancer Institute. The project is also supported by National Institutes of Health intramural funding (to S.M.H). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the National Institutes of Health. The project was also supported by a DoD prostate cancer research program young investigator award (to R.B.) and a computational science and engineering (CSE) fellowship (to J.T.K). Author details Department of Computer Science, University of Illinois at UrbanaChampaign, Urbana, IL 61801, USA. 2Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA. 3Tissue array research program, National Cancer Institute, National Institutes of Health, Bethesda, MD 20850, USA. 4Department of Bioengineering, University of Illinois at Urbana-Champaign, Urbana, IL 61801, USA.
Authors contributions JTK contributed to developing algorithms, programming, and data analysis, and drafted the manuscript. SMH made the diagnoses of tissue samples. SS contributed to developing algorithms and data analysis. RB contributed to the data preparation and data analysis. All authors revised the manuscript and approved the final version. Competing interests The authors declare that they have no competing interests. Received: 21 March 2010 Accepted: 9 February 2011 Published: 9 February 2011 References 1. Jemal A, Siegel R, Ward E, Murray T, Xu JQ, Smigal C, Thun MJ: Cancer statistics, 2006. Ca-a Cancer Journal for Clinicians 2006, 56(2):106-130. 2. Gilbert SM, Cavallo CB, Kahane H, Lowe FC: Evidence suggesting PSA cutpoint of 2.5 ng/mL for prompting prostate biopsy: Review of 36,316 biopsies. Urology 2005, 65(3):549-553. 3. Pinsky PF, Andriole GL, Kramer BS, Hayes RB, Prorok PC, Gohagan JK, P PLCO: Prostate biopsy following a positive screen in the prostate, lung, colorectal and ovarian cancer screening trial. Journal of Urology 2005, 173(3):746-750. 4. Jacobsen SJ, Katusic SK, Bergstralh EJ, Oesterling JE, Ohrt D, Klee GG, Chute CG, Lieber MM: Incidence of Prostate-Cancer Diagnosis in the Eras before and after Serum Prostate-Specific Antigen Testing. Jama-Journal of the American Medical Association 1995, 274(18):1445-1449.

Figure 11 Optimal features for distinguishing cancer and benign tissue samples. The three features shown here are most frequently present in the optimal feature set chosen by the classifier.

Page 15 of 16

12. 13.
Humphrey PA, American Society for Clinical Pathology: Prostate pathology. Chicago: American Society for Clinical Pathology; 2003. Bartels PH, Thompson D, Bartels HG, Montironi R, Scarpelli M, Hamilton PW: Machine vision-based histometry of premalignant and malignant prostatic lesions. Pathol Res Pract 1995, 191(9):935-944. Epstein JI, Netto GJ: Biopsy interpretation of the prostate. Philadelphia: Wolters Kluwer Health/Lippincott Williams & Wilkins, 4 2008. Gleanson DF: Histologic grading and clinical staging of prostate carcinoma. In The Prostate. Edited by: Tannenbaum M. Philadelphia: Lea and Febiger; 1977. Epstein JI, Allsbrook WC, Amin MB, Egevad LL: Update on the Gleason grading system for prostate cancer - Results of an international consensus conference of urologic pathologists. Advances in Anatomic Pathology 2006, 13(1):57-59. Gurcan MN, Boucheron LE, Can A, Madabhushi A, Rajpoot NM, Yener B: Histopathological Image Analysis: A Review. Biomedical Engineering, IEEE Reviews in 2009, 2:147-171. Mulrane L, Rexhepaj E, Penney S, Callanan JJ, Gallagher WM: Automated image analysis in histopathology: a valuable tool in medical diagnostics. Expert Rev Mol Diagn 2008, 8(6):707-725. Madabhushi A: Digital pathology image analysis: opportunities and challenges. Imaging in Medicine 2009, 1(1):7-10. Roula M, Diamond J, Bouridane A, Miller P, Amira A: A multispectral computer vision system for automatic grading of prostatic neoplasia. Biomedical Imaging, 2002 Proceedings 2002 IEEE International Symposium on: 2002 2002, 193-196. Diamond J, Anderson NH, Bartels PH, Montironi R, Hamilton PW: The use of morphological characteristics and texture analysis in the identification of tissue composition in prostatic neoplasia. Human Pathology 2004, 35(9):1121-1131. Stotzka R, Manner R, Bartels PH, Thompson D: A Hybrid Neural and Statistical Classifier System for Histopathologic Grading of Prostatic Lesions. Analytical and Quantitative Cytology and Histology 1995, 17(3):204-218. Wetzel AW, Crowley R, Kim S, Dawson R, Zheng L, Joo YM, Yagi Y, Gilbertson J, Gadd C, Deerfield DW, et al: Evaluation of prostate tumor grades by content-based image retrieval. 1999; Washington, DC, USA SPIE; 1999, 244-252. Smith Y, Zajicek G, Werman M, Pizov G, Sherman Y: Similarity measurement method for the classification of architecturally differentiated images. Computers and Biomedical Research 1999, 32(1):1-12. Jafari-Khouzani K, Soltanian-Zadeh H: Multiwavelet grading of pathological images of prostate. Ieee Transactions on Biomedical Engineering 2003, 50(6):697-704. Farjam R, Slotanian-Zadeh H, Zoroofi RA, Khouzani KJ: Tree-structured grading of pathological images of prostate. Proc SPIE Int Symp Med Imag: 2005; San Diego, CA 2005, 840-851. Doyle S, Hwang M, Shah K, Madabhushi A, Feldman M, Tomaszeweski J: AUTOMATED GRADING OF PROSTATE CANCER USING ARCHITECTURAL AND TEXTURAL IMAGE FEATURES. Biomedical Imaging: From Nano to Macro, 2007 ISBI 2007 4th IEEE International Symposium on: 2007 2007, 1284-1287. Naik S, Doyle S, Feldman M, Tomaszewski J, Madabhushi A: Gland Segmentation and Computerized {G}leason Grading of Prostate Histology by Integrating Low-, High-level and Domain Specific Information. Proceedings of 2nd Workshop on Microsopic Image Analysis with Applications in Biology, Piscataway, NJ, USA: 2007 2007. Tabesh A, Teverovskiy M, Pang HY, Kumar VP, Verbel D, Kotsianti A, Saidi O: Multifeature prostate cancer diagnosis and Gleason grading of histological images. Ieee Transactions on Medical Imaging 2007, 26(10):1366-1378. Huang PW, Lee CH: Automatic Classification for Pathological Prostate Images Based on Fractal Analysis. Ieee Transactions on Medical Imaging 2009, 28(7):1037-1050. Arif M, Rajpoot N: Classification of potential nuclei in prostate histology images using shape manifold learning. Machine Vision, 2007 ICMV 2007 International Conference on: 28-29 Dec 2007, 113-118. Farjam R, Soltanian-Zadeh H, Jafari-Khouzani K, Zoroofi RA: An image analysis approach for automatic malignancy determination of prostate pathological images. Cytometry Part B: Clinical Cytometry 2007, 72B(4):227-240.