Classifying Prostate Cancer Malignancy by Quantitative Histomorphometry

Classifying Prostate Cancer Malignancy by Quantitative Histomorphometry Markus Loeffler,* Lars Greulich, Patrick Scheibe, Philip Kahl, Zaki Shaikhibrahim, Ulf-Dietrich Braumann, Jens-Peer Kuska† and Nicolas Wernert From the Institute for Medical Informatics, Statistics, and Epidemiology (ML, LG, UDB), Interdisciplinary Centre for Bioinformatics (ML, LG, PS, UDB, JPK) and Translational Centre for Regenerative Medicine (PS, JPK), University of Leipzig, Leipzig, Institute of Pathology, Department of Molecular Pathology, University of Bonn (PK, ZS, NW), Bonn and Institute of Pathology, University Hospital of Cologne (PK), Köln, Germany

Purpose: Prostate cancer is routinely graded according to the Gleason grading scheme. This scheme is predominantly based on the textural appearance of aberrant glandular structures. Gleason grade is difficult to standardize and often leads to discussion due to interrater and intrarater disagreement. Thus, we investigated whether digital image based automated quantitative histomorphometry could be used to achieve a more standardized, reproducible classification outcome. Materials and Methods: In a proof of principle study we developed a method to evaluate digitized histological images of single prostate cancer regions in hematoxylin and eosin stained sections. Preprocessed color images were subjected to color deconvolution, followed by the binarization of obtained hematoxylin related image channels. Highlighted neoplastic epithelial gland related objects were morphometrically assessed by a classifier based on 2 calculated quantitative and objective geometric measures, that is inverse solidity and inverse compactness. The procedure was then applied to the prostate cancer probes of 125 patients. Each probe was independently classified for Gleason grade 3, 4 or 5 by an experienced pathologist blinded to image analysis outcome. Results: Together inverse compactness and inverse solidity were adequate discriminatory features for a powerful classifier that distinguished Gleason grade 3 from grade 4/5 histology. The classifier was robust on sensitivity analysis. Conclusions: Results suggest that quantitative and interpretable measures can be obtained from image based analysis, permitting algorithmic differentiation of prostate Gleason grades. The method must be validated in a large independent series of specimens.

Abbreviations and Acronyms GS ⫽ Gleason score PCa ⫽ prostate cancer PSA ⫽ prostate specific antigen ROI ⫽ region of interest Submitted for publication July 14, 2011. Supported by Deutsche Forschungsgemeinschaft Grant WE 1104/11-1) and Deutsche Krebshilfe Grant 107827. * Correspondence: University of Leipzig, IMISE, Härtelstr. 16-18, D-04107 Leipzig, Germany (telephone: ⫹49-341-97-16100; e-mail: loeffler@imise. uni-leipzig.de). † Dedicated to our colleague J.P. Kuska, who passed away on July 1, 2009. We admire his intellect and encouragement.

Key Words: prostate; prostatic neoplasms; neoplasm grading; image processing, computer-assisted; pathology PROSTATE cancer is among the most common cancers in Western countries. Thus, considerable effort has been made to unravel its development and progression.1 The tumor shows considerable heterogeneity not only in molecular terms but also in malignancy grade,2 clinical behavior and outcome.1,3

Systematic PSA screening has led to increasing PCa detection.4 However, the precise clinical course of a specific tumor is still difficult to identify while most efforts have been aimed at molecular pathways, gene fusion5 and microRNA.6 – 8 However, to date pretherapy PSA, clinical stage and particularly histological malig-

0022-5347/12/1875-1867/0 THE JOURNAL OF UROLOGY® © 2012 by AMERICAN UROLOGICAL ASSOCIATION EDUCATION

Vol. 187, 1867-1875, May 2012 Printed in U.S.A. DOI:10.1016/j.juro.2011.12.054

AND

RESEARCH, INC.

www.jurology.com

1867

1868

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

Table 1. Histopathological results of probe by Gleason grade, tumor stage and lymph node status in 125 patients No. Pts Gleason grade: 3 4 5 pT stage: 1 2 3 4 pN stage: 0 1

74 45 6 1 72 46 6 115 10

nancy grade evaluated by pathologists in prostate core biopsies have been the main factors guiding treatment strategy.9 Since the consensus conference in 1993, GS10 has been internationally recommended as the most suitable parameter to determine the malignant potential of PCa. The system relies on qualitative analysis of architectural features of cancer tissue. Most clinical studies are now based on this score. Other grading systems exist, including those of WHO and Mostofi,11 and Helpap et al,12 which evaluate nuclear atypia as well as histological patterns. However, intra-observer and interobserver reproducibility remains a problem inherent to all grading systems, besides the representativeness of prostate core biopsies.13 The 2005 revision of the Gleason scoring system ameliorated interobserver reproducibility and correlation of findings in biopsies and prostatectomy specimens.14,15 Several groups have proposed quantitative algorithms for PCa diagnosis and Gleason grading using digitized microscopic images.16 –19 The typically achieved accuracy for classification into low and high grade classes was 77.3% and 81.0%, respectively, in 2 independent studies.16,19 However, to date none of these methods has attained routine usability for histopathological diagnosis, partly due to insufficient discriminatory features of the classifiers proposed in these studies. Previously we used image processing to quantify malignant changes in prostate tissue sections immunohistochemically stained for PSA.20 A quantitative measure calculated using the inverse solidity of PSA positive prostate tissue segments correlated highly with malignancy grade according to Gleason14 and Helpap et al.12 These previous findings led us to perform more extended digital image analysis in this study by applying histomorphometry on routine hematoxylin and eosin stained sections with refinement of morphometric measures for PCa growth patterns in a large series of probes.

MATERIALS AND METHODS Tissue Probe Selection We used conventional hematoxylin and eosin stained sections obtained for routine histopathological diagnostic assessment from the prostatectomy specimens of 125 patients treated for PCa between 1996 and 2004 at the Department of Urology, University of Bonn. Mean patient age was 66.2 years (median 66.1, range 49 to 78). No patient received hormone ablative or other conservative treatment preoperatively. Mean PSA at diagnosis was 17.6 ng/ml (median 11.06, range 3.15 to 75.92). Cancer stage was predominantly pT2 and pT3. Table 1 lists further grade and stage related patient characteristics. Slides with hematoxylin and eosin stained sections were captured with a nominal squared pixel size of 0.332 ␮m2 using a MIRAX MIDI slide scanner (Carl Zeiss®) (fig. 1, A). Analysis was done within ROIs where carcinoma was detected. ROIs were defined by a pathologist

Figure 1. Chain of applied image processing steps, including prostate histological section image with PCa including defined ROI (rectangle) (A), edge preserving smoothing by total variation filtering ROI of 1,243 ⫻ 2,052 ␮m enlarged (B), color deconvolution applied to ROI with light blue areas indicating hematoxylin and, thus, enhancing all glandular structures (C) and ROI after simple thresholding (D), providing binary image segmentation of glandular objects.

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

(NW), who performed Gleason grading only for these ROIs, deliberately disregarding other regions or full section related Gleason scoring. Those who performed image analysis were blinded to grade. If multiple PCa ROIs were defined per section, only the largest ROIs of those graded highest were considered. Of 125 patient probes Gleason grade 3, 4 and 5 disease was found in 74, 45 and 6, respectively.

Digital Imaging Processing. ROIs with a rectangular edge length of between 1.01 and 3.38 mm were coarsened (effective pixel size 1.322 ␮m2 with no subcellular resolution), resulting in an image raster size of between 768 and 2,560 pixels. To effectively decrease image inhomogeneity we applied total variation filtering,21 which provided edge preserving smoothing as a preprocessing step to facilitate subsequent automatic prostate gland segmentation (fig. 1, B). Since prostate neoplastic gland segmentation was the prerequisite for subsequent histomorphometry, it was implemented using the color deconvolution approach by refer-

1869

ring to a published red-green-blue color vector for hematoxylin22 to provide a pixel-wise hematoxylin intensity estimate (fig. 1, C). In this study we interactively selected binarization thresholds for the neoplastic gland segmentation of each ROI image to search for optimum visual agreement between glandular structures in color and in binary images (figs. 1, C and D, and 2). Depending on staining and signal intensity in the red-green-blue color space we noted individual optimum threshold values of 15 to 100 U on the 8-bit gray value scale (offset ⫾ 0) (fig. 2). Interactive visual 1-by-1 comparisons of binary vs original hematoxylin and eosin stained images led to rather unambiguous individual threshold choices (figs. 1, B to D and 2). Figure 3 shows a histogram of all ROI threshold values for neoplastic glandular epithelium segmentation that was obtained by visual inspection. The histogram particularly reflects specimen related staining variability. Since variation of the threshold affects the resulting binary image, on systematic sensitivity analysis we also

Figure 2. Segmentation results when manually optimized thresholds are varied show specific ROIs as stained image with respective binarization (bin.) results (A to D). Shifting ⫺20 in tendency led to glandular structure over segmentation but 10 provided some under segmentation. Physical ROI size was 2,703 ⫻ 1,687 (A), 2,028 ⫻ 1,352 (B), 2,041 ⫻ 1,655 (C) and 2,185 ⫻ 2,013 (D) ␮m. Offset of ⫺20 (asterisk) attained limitation of lowest permissible threshold 10 (C). Concave margins were due to previous core biopsies punched from paraffin block (B).

1870

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

optimum visual agreement threshold (offset ⫾ 0). To avoid thresholds that were too extreme permissible offset variations were limited by introducing minimum and maximum thresholds of 10 and 150, respectively. Spot tests did not document significant hematoxylin related intensity below or above these values. To further analyze robustness we chose a random number from the interval ⫺20 to 20 and added it to the optimum visual agreement thresholds. Thus, together with the mentioned 5 neoplastic glandular epithelium segmentation thresholds we generated 6 binary image data sets per ROI. Figure 2 shows parts of this variation for 4 ROIs.

Figure 3. Histogram plot shows achieved optimum visual agreement thresholds used for binarization of 125 ROI images.

investigated the extent to which the choice of binarization threshold affected our morphometric analysis. For each ROI image we added certain offsets of ⫺20, ⫺10, 10 and 20 U on an 8-bit gray value scale to the previously found

Analysis. For each ROI binary segmentation images were taken for morphological parameter calculations. Respective quantitative histomorphometric analysis on segmented neoplastic glandular epithelium objects (sets of connected foreground pixels with a minimum area of 1,000 pixels, ie 1,742 ␮m2) were done using the computer algebra system Mathematica 7 (Wolfram Research, Champaign, Illinois) together with the associated Digital Images package developed by one of us (JPK). To quantify these glandular objects we applied 2 geometric approaches for descriptive histomorphometric measures, including inverse solidity (S) and inverse compactness (C).

Histomorphometric Measures Inverse solidity. Figure 4, A shows the scheme illustrating inverse solidity. The black segment was considered 1 object representing 2 adjacent neoplastic prostate glandu-

Figure 4. Schematic visualization shows 2 morphometric measures from sample gland (orange object near center, B) (A). Blue outline indicates area inside convex hull (Ah). Orange outline indicates object perimeter (Po). Orange circle indicates circle with same perimeter (Po). Glandular object area (Ao) corresponds to number of black pixels. Pseudo color representation shows 2 ROIs with Gleason grade 3 (B) and 4 (C), respectively. Objects used for morphometric analysis were drawn in pseudo colors. Gray objects indicate small parts that were sorted out by algorithm. Physical extent is 1,243 ⫻ 2,052 (fig. 1, B) and 2,365 ⫻ 2,703 ␮m.

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

lar structures. Inverse solidity is defined as the ratio of the area enclosed by the convex hull (Ah) and the area covered by the object (Ao), that is equation 1, So ⫽ Ah/Ao. Filled convex objects shaped as a circle, ellipse or rectangle show a minimum inverse solidity of So min ⫽ 1. Annulus-shaped, folded or concave structures have an inverse solidity of larger than 1. Since we aimed to obtain a histomorphometric quantification representing many objects in an ROI, we used an object area (Ao) normalized average by determining weighted inverse solidity (Sw) using equation 2, Sw ⫽ 兺Soi·Aoi ⁄ 兺Aoi. Inverse compactness. Figure 4, A shows the scheme illustrating inverse compactness (C). The perimeter of the same glandular object (Po) was assigned a circle with a diameter (d) and an identical perimeter (Po ⫽ Pc ⫽ ␲dc) covering a circular area of Ac ⫽ ␲d2c ⁄4 ⫽ P2o ⁄ 共4␲兲. Inverse compactness was defined as the ratio of the respective circular area and the area of the object as (equation 3) Co ⫽ Ac/Ao, which can also be expressed as (equation 4) Co ⫽ P2o ⁄ 4␲Ao. If an object occurred as a circular area without an inner lumen, it would show a minimum inverse compactness of Co min ⫽ 1. For neoplastic prostate glands typical Co values are 10 or greater. Obviously the more lobulated a glandular object, the larger its perimeter and, thus, the larger the Co. Analogous to equation 2, we determined the object area (Ao) normalized, ROI related average providing weighted inverse compactness (Cw) using equation 5, Cw ⫽ 兺Coi·Aoi ⁄ 兺Aoi. Figure 4, B and C shows 2 PCa examples with differently shaped glandular objects. Figure 4, B shows an ROI representing a Gleason grade 3 tumor (99 objects) with rather low weighted inverse compactness (Cw ⫽ 23). Figure 4, C shows an ROI of a Gleason grade 4 tumor (16 objects) with Cw ⫽ 612. Weighted inverse solidity values were Sw ⫽ 2.20 and 1.76 for the depicted grade 3 and 4 ROIs, respectively. Clearly inverse solidity (Sw) and inverse compactness (Cw) measure different properties. A simple glandular tubule structure more or less appears as an annulus with Cw in the low 2-digit range and an Sw of 1.2 to 3 depending on the respective lumen size. A densely packed ROI covering an invasive PCa might contain many lobulated interconnected grape-like structures in a small area, leading to a high Cw and a rather small Sw. On the other hand, a glandular structure with many separated perforated objects would show a high Sw and a rather low Cw. Thus, a combination of the 2 measures appears promising to differentiate glandular patterns.

RESULTS Univariate Analysis Figure 5, A shows box and whisker plots summarizing measurements of the inverse compactness (Cw) with regard to binarization threshold offsets ⫾ 0, ⫺20 and 10 on the 8-bit gray value scale. Cw values spread over more than 2 orders of magnitude showed large morphological shape differences. Probes classified as Gleason grade 3 had approximately Thirtyfold lower in-

1871

verse compactness than probes with a Gleason grade of 4 or 5. Even when we varied the binarization threshold, differences remained strong. Figure 5, B shows that the inverse solidity (Sw) morphometric measure varied reciprocally by a factor of about 2. Smaller Sw values seemed to be associated with higher Gleason grade. Unlike Cw results, Sw measurements for Gleason grade 4/5 appeared to depend more on the binarization threshold. Multivariate Analysis and Morphometric Gleason Classifier Multivariate analysis revealed that the 2 morphometric measures inverse compactness (Cw) and inverse solidity (Sw) carried information useful for constructing a morphometry based Gleason grade classifier. Figure 5, C shows scatterplots illustrating the combination of the 2 measures. Two nearly distinct clusters emerged. Gleason grade 3 tended to have lower inverse compactness and slightly increased inverse solidity than Gleason grade 4/5 probes. Multiple logistic regression models were used to determine whether and how well Gleason grade could be predicted from logarithmized inverse compactness and inverse solidity values alone (table 2). Model 1 shows results using the reference (optimum visual agreement) binarization threshold, offset ⫾ 0. For Gleason grade 3 f(z) approached 0 while for Gleason grades 4/5 it was close to 1. Six regression models were obtained, which differed in the applied binarization threshold offset. Under common p value conventions regression coefficients were highly significant, showing that the separation worked well. We also constructed contour lines for the probability of belonging to Gleason grades 4/5 (fig. 5, C). For the 50% line we obtained a correct classification of 94.4%. We found more than 90.4% of all observations when focusing only on regions outside the 20% to 80% zone. For these cases the misclassification rate decreased from 5.6% to 3.5% (table 2). To investigate the sensitivity or vice versa the robustness of the binarization threshold choice we constructed analogous models for varied thresholds with systematic (models 2 to 5) or random (model 6) offset values. Table 2 shows the robustness of the results to these variations, leading to similar misclassification results. Based on the correct classification rate lowering the threshold (models 2 and 3) potentially risked some glandular structure over segmentation, which would not impair the overall performance of this model type. Raising this threshold led to under segmentation (models 4 and 5) but clearly decreased the correct classification rate. This was also observed for (restricted) random threshold modifications (model 6). Regression analysis was done using SPSS® 18.

1872

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

Figure 5. Univariate analysis of obtained morphometric measurements for Gleason grade 3 vs 4/5 (A and B). Box and whisker plots show logarithmized inverse compactness (log10 Cw) of 3 binarization threshold offsets (A). Larger values seemed associated with higher Gleason grade. Circles represent mild outliers, defined as distance less threefold IQR, ie box length. Stars represent extreme outliers. Inverse solidity (Sw) appeared lower for higher Gleason grades since higher de-differentiation led to more densely packed glandular tissue (B). Over segmentation (threshold offset: ⫺20) decreased inverse solidity. Scatterplots demonstrate all measured regions with inverse solidity (Sw) plotted against logarithmized inverse compactness (log10 Cw) (C). Linear classification function was obtained from logistic regression models. Probability of Gleason grade 3 tumor exceeded 80% if localized in upper left corner. Probability of Gleason grade 4 or 5 tumor exceeded 80% if localized in lower right corner. Few cases were noted between 20% and 80% percentile estimates (outer lines) (table 2). Middle line indicates 50% percentile.

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

1873

Table 2. Regression model using basic logistic form f (z) ⫽ 1/(1 ⫹ e⫺z) with z ⫽ ␤0 ⫹ ␤1log10Cw ⫹ ␤2Sw to assess whether Gleason grade could be predicted from combining logarithmized inverse compactness log10Cw and inverse solidity Sw Logistic Regression

% Outside 20% to 80% Zone

Model

Binarization Threshold Offset

Intercept ␤0

Coefficient ␤1(log10Cw)*

Coefficient ␤2(Sw)

% Correct Classification

Observation

Correct Classification

1 2 3 4 5 6

0 ⫺10 ⫺20 10 20 20–⫺20

⫺1.088 3.445 6.224 ⫺0.862 ⫺0.038 ⫺1.596

4.444 3.868 3.423 4.668 4.949 3.225

⫺4.334† ⫺6.764* ⫺8.083* ⫺3.603† ⫺3.678† ⫺2.454†

94.4 95.2 95.2 88.8 87.2 87.2

90.4 95.2 92.8 80.0 69.6 81.6

96.5 96.6 97.4 95.0 90.8 92.2

* p ⬍0.001. † p ⬍0.05.

DISCUSSION PCa is a heterogeneous disease1 and Gleason scoring is considered a major prognostic factor for this tumor.14,23 When using the revised version for Gleason scoring, estimated interobserver reproducibility is greater than 80%.24 However, reproducibility remains critical in all grading systems due to subjective semiquantitative evaluation of slides and different degrees of clinical pathologist training. Thus, a more objective, quantitative assessment of cancer dedifferentiation patterns is desirable. We report the possibility of designing a reliable algorithmic histomorphometric classifier to grade PCa by exploiting 2 novel measures. The classifier requires calculating inverse solidity and inverse compactness from gland segmented binarized images. When evaluated together, the 2 morphometric measures effectively discriminated between ROIs with Gleason grade 3 and 4/5. However, the classifier did not provide a distinction between grades 4 and 5, probably since they have only slight differences, which may be difficult to assess even through direct visualization by a uropathologist. Also, we did not have enough specimens to reliably assess the usefulness of our method to distinguish between grades 4 and 5. The clinical pathologist cannot be replaced by our approach since the pathologist must diagnose prostate carcinoma. This diagnosis includes architectural features that may be assessed by our quantitative approach as well as further criteria, particularly nuclear atypia and absent basal cells, which often require immunohistochemistry. Thus, our approach could at best help classify lesions of uncertain malignancy. Our proof of principle study was applied to single ROIs of prostate carcinoma sections. We did not assess multiple ROIs in the same specimen. However, we expect our data to be reproducible in multiple ROIs in the same specimens since we achieved comparable results in similar ROIs in a large cohort of 125 patients. Our method uses a segmentation step that is implemented as manual thresholding under visual in-

spection of the obtained binarizations. Figure 3 shows that the probes led to great heterogeneity of the respective binarization thresholds, primarily due to variations in hematoxylin and eosin staining intensity and color expression. Another factor may be the retrospective use of slides from an 8-year period. Future study will clarify how to introduce automatic prostatic gland segmentation. We rereviewed all cases with a discordant assignment of Gleason grade using a pathologist and our algorithm. We found 2 common features for all discordant cases. The first feature, which concerned most ROIs, consisted of fusing glandular structures, for which it is difficult to assign a Gleason grade of 3 or 4 even by a uropathologist. The second feature, which concerned only a few ROIs, consisted of nodular cribriform formations, for which it is easy to assign a Gleason grade 4 visually. In contrast, this rare particular pattern is treated like glandular structures by our digital algorithm, which might be ameliorated for this pattern. When visually reevaluating a gold standard, rereviewing assigned an alternate grade to only about 50% of all discordant ROIs. Thus, these difficulties are inherent to the Gleason grading system. Although our method cannot yet be applied for online routine diagnosis, we believe that it can be used in a setting of reference pathology assessments, eg in clinical trials, in which it would be possible to process a large series of images off line for scientific evaluation. Our approach could also be applied to the WHO11 and Helpap et al12,25 grading systems, and together with the quantitative evaluation of nuclear atypia developed in a previous study.26 Notably it was not our objective to consider prognosis, which in turn would require at least 2 or several ROIs. Ideally whole tumor sections would be assessed and a joint computed morphometric score would be constructed. This is feasible and would enhance the benefits of our system. Results would have to be compared to the long-term outcome. Moreover, prognostic evaluation would also warrant a sufficiently large cohort of cases discordantly classified by the algorithm and the reference pathologists, and such a data set

1874

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

was not available. This approach would also enable us to decide whether our system of Gleason grading could prove useful to train young or experienced pathologists in Gleason scoring. For comparative purposes we referred to previous studies of quantitative algorithms for computer assisted diagnosis with regard to Gleason grading. Stotzka et al applied a neural network to statistical features of spatial nuclei distribution and distinguished moderately vs poorly differentiated samples with 77.3% accuracy.16 Diamond et al used morphometric and texture features to separate stroma from normal and cancerous image regions with less than 80% accuracy.18 Tabesh et al combined features from color, fractal dimensions, texture and some morphometric clues and achieved at best 81% accuracy to discriminate low (2/3) from high (4/5) grade Gleason scores.19 Jafari-Khouzani and Soltanian-Zadeh selected the most discriminative image energy and entropy features by simulated annealing.17 Using a k-nearest neighbor classifier they typically achieved 90% grading accuracy. Huang and Lee applied a sophisticated classifier using 8 fractal dimension related feature sets to discriminate 4 Gleason grades, achieving around 90% accuracy.27 Xu et al referred to needle core biopsies and used an active contour model for glandular inner lumen segmentation and medial axis shape model comparison.28 Although they achieved excellent gland segmentation results, the differentiation accuracy between Gleason grade 3/4 and benign referents remained at less than 83%.

CONCLUSIONS The major result of our analysis is that 2 simple, intuitively understandable measures referring to

the size and contour of segmented, gland related objects seem to enable good, reliable discrimination of low and high Gleason grades. We believe that our analysis leads the way to a formalized, standardized digital morphometry of PCa sections. Scanning whole histological slides and treating huge data quantities in digitized images is becoming increasingly feasible and rapid. Thus, suitable mathematical morphometric measures of cancer growth patterns, such as inverse solidity and inverse compactness, may be systematically investigated in large patient cohorts and become novel quantitative pathological measures capable of contributing to the prediction of cancer aggressiveness and prognosis. In future studies our method must be scaled up and extended to full morphometric scoring using information from several simultaneous core biopsies or complete histological slides. Since we can quantify the contribution of grade 3 and 4 regions over all slides, this should also permit more objective subdivision of GS 7 cases into 7a (GS 3 ⫹ 4) and 7b (GS 4 ⫹ 3), which differ in prognosis.15

ACKNOWLEDGMENTS Nico Scherf, Institute of Medical Informatics and Biometry, Dresden University of Technology; and Andreas Heffel, Interdisciplinary Center for Bioinformatics, Karsten Winter, Translational Center for Regenerative Medicine and Priya Pathak, Center for Biotechnology and Biomedicine, University of Leipzig, assisted with discussion. Jacqueline Czerwitzki, Michaela Cuschie and Susanne Steiner, Institute of Pathology, University of Bonn, provided technical assistance.

REFERENCES 1. Mackinnon AC, Yan BC, Joseph LJ et al: Molecular biology underlying the clinical heterogeneity of prostate cancer: an update. Arch Pathol Lab Med 2009; 133: 1033.

6. Wegiel B, Evans S, Hellsten R et al: Molecular pathways in the progression of hormone-independent and metastatic prostate cancer. Curr Cancer Drug Targets 2010; 10: 392.

2. Arora R, Koch MO, Eble JN et al: Heterogeneity of Gleason grade in multifocal adenocarcinoma of the prostate. Cancer 2004; 100: 2362.

7. Pang Y, Young CY and Yuan H: MicroRNAs and prostate cancer. Acta Biochim Biophys Sin (Shanghai) 2010; 42: 363.

3. Dall’Era MA and Kane CJ: Watchful waiting versus active surveillance: appropriate patient selection. Curr Urol Rep 2008; 9: 211.

8. Gandellini P, Folini M and Zaffaroni N: Emerging role of microRNAs in prostate cancer: implications for personalized medicine. Discov Med 2010; 9: 212.

4. Mistry K and Cable G: Meta-analysis of prostate-specific antigen and digital rectal examination as screening tests for prostate carcinoma. J Am Board Fam Pract 2003; 16: 95.

9. Roach M III, Small E, Reese DM et al: Urologic and male genital cancers: prostate cancer. In: Clinical Oncology: A Multidisciplinary Approach for Physicians and Students, 8th ed. Edited by P Rubin. Philadelphia: WB Saunders 2001; pp 538 –550.

5. Narod SA, Seth A and Nam R: Fusion in the ETS gene family and prostate cancer. Br J Cancer 2008; 99: 847.

10. Gleason DF: Classification of prostatic carcinomas. Cancer Chemother Rep 1966; 50: 125.

11. Mostofi FK, Sesterhenn IA, Davis CJ et al: Histological Typing of Prostate Tumours. Berlin: Springer Verlag 2002; pp 15–16. 12. Helpap B, Böcking A, Dhom G et al: Classification, histologic and cytologic grading and regression grading of prostate cancer. Urologe A 1985; 24: 156. 13. Burchardt M, Engers R, Müller M et al: Interobserver reproducibility of Gleason grading: evaluation using prostate cancer tissue microarrays. J Cancer Res Clin Oncol 2008; 134: 1071. 14. Epstein JI, Allsbrook WC Jr, Amin MB et al: The 2005 International Society of Urological Pathology (ISUP) consensus conference on Gleason grading of prostatic carcinoma. Am J Surg Pathol 2005; 29: 1228. 15. Helpap B and Egevad L: Modified Gleason grading. An updated review. Histol Histopathol 2009; 24: 661.

CLASSIFYING PROSTATE CANCER MALIGNANCY BY QUANTITATIVE HISTOMORPHOMETRY

16. Stotzka R, Manner R, Bartels PH et al: A hybrid neural and statistical classifier system for histopathologic grading of prostatic lesions. Anal Quant Cytol Histol 1995; 17: 204.

In: Bildverarbeitung für die Medizin 2008 —Algorithmen, Systeme, Anwendungen. Edited by T Tolxdorff, J Braun, TM Deserno et al. Berlin: Springer-Verlag 2008; pp 358 –362.

17. Jafari-Khouzani K and Soltanian-Zadeh H: Multiwavelet grading of pathological images of prostate. IEEE Trans Biomed Eng 2003; 50: 697.

21. Chan TF, Osher S and Shen J: The digital TV filter and nonlinear denoising. IEEE Trans Image Process 2001; 10: 231.

18. Diamond J, Anderson NH, Bartels PH et al: The use of morphological characteristics and texture analysis in the identification of tissue composition in prostatic neoplasia. Hum Pathol 2004; 35: 1121.

22. Ruifrok AC and Johnston DA: Quantification of histochemical staining by color deconvolution. Anal Quant Cytol Histol 2001; 23: 291.

19. Tabesh A, Teverovskiy M, Pang HY et al: Multifeature prostate cancer diagnosis and Gleason grading of histological images. IEEE Trans Med Imaging 2007; 26: 1366. 20. Braumann UD, Kuska JP, Löffler M et al: Quantify prostate cancer by automated histomorphometry.

23. Amin M, Boccon-Gibod L, Egevad L et al: Prognostic and predictive factors and reporting of prostate carcinoma in prostate needle biopsy specimens. Scand J Urol Nephrol Suppl 2005; p 20. 24. Melia J, Moseley R, Ball RY et al: A UK-based investigation of inter- and intra-observer reproducibility of Gleason grading of prostatic biopsies. Histopathology 2006; 48: 644.

1875

25. Helpap B: Review of the morphology of prostatic carcinoma with special emphasis on subgrading and prognosis. J Urol Pathol 1993; 1: 3. 26. Makarov DV, Marlow C, Epstein JI et al: Using nuclear morphometry to predict the need for treatment among men with low grade, low stage prostate cancer enrolled in a program of expectant management with curative intent. Prostate 2008; 68: 183. 27. Huang PW and Lee CH: Automatic classification for pathological prostate images based on fractal analysis. IEEE Trans Med Imaging 2009; 28: 1037. 28. Xu J, Sparks R, Janowczyk A et al: High-throughput prostate cancer gland detection, segmentation, and classification from digitized needle core biopsies. Prostate Cancer Imaging 2010; LNCS 6367: 77.