[945] Quantitative Digital Analysis of Histologic Features of Prostatic Adenocarcinoma

Y Peng, Y Jiang, L Eisengart, S-T Chuang, FH Straus, XJ Yang. University of Chicago, Chicago, IL; Northwestern University, Chicago, IL

Background: Accurate diagnosis of prostatic adenocarcinoma in clinical practice is based on a constellation of histologic criteria, which can be subjective and variable. We aim to develop a computer method to aid the diagnosis of prostatic adenocarcinoma. We quantitatively analyzed several important histologic features of the prostate facilitated by a computer method.
Design: We used a scope-mount digital camera to collect digital bright-field histologic images of prostatectomy specimens from two institutions, and obtained a total of 173 images at 100X (64 malignant, Gleason 3, and 109 benign) and 142 images at 200X (72 malignant, Gleason 3-5, and 70 benign). We developed computer methods to identify individual glands and glandular lumens from 100X images, and nucleoli from 200X images. Five histologic features extracted from 100X images (Figure) were combined into a single score using a linear classifier. We evaluated the features using area under the receiver operating characteristic (ROC) curve (AUC).
Results: All features were effective in separating malignant (adenocarcinoma) from benign glands (Figure). Linear combinations of the five features extracted from 100X images produced AUC values of 0.92 to 0.96 (1.0 indicates perfect diagnosis; 0.5 indicates random call). The features extracted from 100X images show moderate to strong correlation (Table).
Conclusions: Quantitative measurement of prostatic glandular size, luminal size, lumen circularity, gland- and lumen-to-stroma ratios, and nucleolar size, facilitated by a computer segmentation technique that we have developed, are concordant with pathologic diagnosis. These results suggest potential usefulness of quantitative digital analysis of histologic features of the prostate in surgical specimens to provide an ancillary tool for improving the diagnosis of prostate cancer.

Table: Pearson Correlation Coefficients Between Features Extracted from 100X Images
Lumen sizeLumen circularityGland-stroma ratioLumen-stroma ratio
Gland size0.70-0.480.610.55
Lumen size-0.450.520.68
Lumen circularity-0.56-0.29
Gland-stroma ratio0.53





Category: Genitourinary (including renal tumors)

Tuesday, March 23, 2010 1:00 PM

Poster Session IV # 122, Tuesday Afternoon

 

Close Window