Identifying pathological groups from MRI in prostate cancer using graph representation learning

Abstract

Multiparametric magnetic resonance imaging (mpMRI) plays a critical role in prostate cancer (PCa) diagnosis, aiding in clinical trial evaluation and personalized treatment planning. We propose a novel prediction approach based on graph representation learning (GRL) integrating mpMRI images for classifying International Society of Urological Pathology (ISUP) grade groups. We initially constructed a mpMRI image graph representation learning (MMIGRL) model using a graph-based representation fusion approach. After constructing a heterogeneous graph integrating T2-weighted and apparent diffusion coefficient (ADC) images, we applied an adaptive method that dynamically adjusted the edges within the graph to reveal relationships within the mpMRI images. Subsequently, we developed a residual graph convolution network (Res-GCN) model to establish the mapping relationship between the mpMRI fusion graph and ISUP grade groups. The results of 5-fold cross-validation demonstrated excellent overall performance, and the prediction results using independent testing datasets confirmed the model's generalizability. The results of ablation studies highlighted the positive impact of integrating T2-weighted and ADC images and the effectiveness of residual connections in learning hierarchical features within the graph structure. The results of comparative experiments validated the superiority of the proposed MMIGRL model compared with convolutional neural network (CNN) and vision transformer (ViT) models. This approach holds promise for accurate ISUP grade group classification in PCa diagnosis, contributing to improved patient care and clinical decision making.