Computer extracted nuclear features from tumor and benign regions of Feulgen and H&E images to help predict recurrence in prostate cancer patients following radical prostatectomy.

Abstract

e16556 Background: Following radical prostatectomy, around 30% of prostate cancer (PCa) patients experience biochemical recurrence (BCR). H&E highlights nuclear morphology and Feulgen reflects nuclear DNA content, a feature linked to PCa presence and aggressiveness. In this work we sought to explore whether computer extracted measurements of tumor morphology and tumor adjacent benign regions on H&E and Feulgen tissue images could predict BCR. Methods: We used 108 patients (59 BCR and 49 non-recurrence (NR)) and each patient had 242 QH features calculated from both the tumor and benign region of stained TMA core images. Feature selection was performed on a training set (30 BCR, 24 NR) to select the 10 most discriminating tumor and tumor adjacent benign features of each stain. A random forest classifier was trained with features so identified and validated on a test set (29 BCR, 25 NR) to predict BCR. Predictions were displayed using Kaplan-Meier analysis and area under the ROC curve (AUC). Results: The most discriminating feature from the tumor regions of the H&E stain was Fourier descriptors of nuclear shape and from the Feulgen stain was texture intensity while from the benign regions it was invariant moments of nuclear shape and texture contrast energy. Combining the significant features from tumor and tumor adjacent benign regions from H&E and Feulgen resulted in the highest accuracy and a statistically significant difference (p < 0.05) via a log-rank test (Table 1). Gleason score did not show statistically significant differences and had the lowest AUC. Conclusions: Combining nuclear morphology and DNA related features of the tumor and tumor adjacent benign regions enabled accurate prediction of BCR. With additional multi-site validation, the combined H&E + Feulgen classifier could allow better risk stratification and post-surgical patient management. [Table: see text]