Classification before Segmentation: Improved U-Net Prostate Segmentation

Prostate segmentation is a necessary pre-processing step for computer-aided detection and diagnosis algorithms for prostate disorders and associated cancers. Deep learning models like U-Net offer the potential for performing classification and segmentation in a single step. We evaluated the U-Net model for prostate segmentation on 1,235 Magnetic Resonance (MR) images from 39 patients with prostate cancer. On our data set, the U-Net models generated a substantial number of false-positive predictions (average precision of 67.6%). We propose separating classification and segmentation into distinct tasks implemented via a pipeline. Our classifier achieved an average ROC AUC of 97.7% (SD 1.0%) in four-fold cross-fold validation. Compared to the U-Net model alone, our pipeline increased overall agreement between the predicted and human-annotated masks (Dice score of 0.907 vs 0.758 and precision of 83.2% vs 67.6%) without a significant loss in recall (83.4% vs 85.6%). Performing classification and segmentation in a single step is a desirable goal. Our results suggest, however, that improvements are still needed before U-Net offers comparable precision to using a separate classifier.