A Deep Learning Framework Predicts Rectal Cancer Treatment Response via Pretreatment 3D Transrectal Ultrasound: A Multicenter Study

Abstract

Accurate prediction of pathological complete response (pCR) following neoadjuvant therapy is crucial for personalized treatment planning in patients with locally advanced rectal cancer (LARC). This study aimed to explore the application of 3D transrectal ultrasound (TRUS) for this purpose. In this study, 538 LARC patients from five hospitals were enrolled and divided into training (n = 348), internal validation (n = 87), and external validation (n = 103) cohorts. Using pretreatment 3D TRUS data of the rectal tumors, a deep learning framework comprising an automated segmentation model and a pCR prediction model was constructed and validated. The segmentation model achieved an excellent Dice score of 0.89. The predictive model, when trained on data sampled at 9° or 18° intervals, yielded area under the curve (AUC) values of 0.92 (internal validation) and 0.85–0.86 (external validation), with accuracies of 86.2%–90.8% and 81.6%–83.5%, respectively; decision‐curve analysis also demonstrated clinically meaningful net benefits. This study presents the first deep learning framework based on 3D TRUS for pCR prediction in LARC patients, offering an automated and reproducible tool that may support clinical decision‐making in rectal cancer management. However, further validation in larger, multi‐device cohorts is essential before clinical application.