PROper-Net: A Deep-Learning Approach for Prostate’s Peripheral Zone Segmentation based on MR imaging

Prostate cancer detection and characterization on Magnetic Resonance Images (MRI) requires accurate segmentation of the prostate gland and the prostatic sub-regions. With the majority of tumoral lesions located in the prostate’s peripheral zone, a precise segmentation of this region is imperative for tumor characterization. Despite the growing success of Convolution neural networks (CNN) in the task of prostate gland segmentation, there a is a knowledge gap in the performance of such networks for segmenting prostatic subregions. In the present work, we propose an novel Deep Learning (DL) approach, named PROper-Net, for segmenting the prostate’s peripheral zone on T2-weighted (T2w) MR images. Our network was compared against four state-of-the-art encoder–decoder CNNs: the original Unet, and its extensions Unet++, Unet3+, and Bridged Unet. Overlap- and distance-based metrics were computed to assess models’ performance and to quantify the superiority of the proposed segmentation approach. The results show that the proposed network successfully outperforms existing networks for the peripheral zone segmentation task, yielding a median performance of 0.74 in terms of Dice Score and 0.88 in terms of balanced accuracy. The improvement in segmentation performance was significant (p-value 0.05) with respect to Unet, Unet++, Unet3+ for all the evaluation metrics while for Bridged Unet significant improvement was achieved for Dice Score, Balanced Accuracy, Sensitivity, and Rand Error Index.