3D convolutional network with edge detection for prostate gland and tumor segmentation on T2WI and ADC

Abstract

Prostate cancer (PCa) ranks as the third most prevalent cancer globally. Accurate segmentation of the prostate and tumor in magnetic resonance image (MRI) has vital importance for supporting auxiliary diagnosis and subsequent research on prostate diseases. Specifically, T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) sequences serve as crucial references for identifying the prostate boundaries and locating the lesion areas. In this study, we propose a multi-segmentation model that takes T2WI and ADC and outputs the corresponding prostate gland and tumor segmentations. In the network, we introduced SE-Net to enhance feature processing with the combination of Cased Pyramid Convolution and Attention Gate. We utilized convolutions with different scales to implement residual connections for dealing with the inconsistent spatial information of the features and optimizing the training process. One major challenge in prostate segmentation arises from the blurred image of the tissue edge regions. To overcome this difficulty, we designed a pure convolution-based edge detection block in the model's architecture to constrain edge regions simultaneously. This approach improved the segmentation accuracy of the indistinct boundaries between the prostate gland, tumor, and surrounding anatomical structures. Upon evaluating with magnetic resonance images from Guizhou Provincial People's Hospital, we achieved the best scores in two metrics: Average boundary distance, and Hausdorff distance for both the prostate gland and the tumor segmentation results. The external validation results demonstrated the model's generalization ability, substantiating its efficacy when applied to data from other sources.