Application of deep learning on MRI for prognostic prediction in rectal cancer.

objectivesPretreatment MRI was employed to develop and validate a combined model integrating clinical features with deep learning for rectal cancer. MATERIALS AND

methodsWe retrospectively collected 458 patients from three hospitals and followed them up for at least 3 years. Clinical, pathological and imaging data were collected. Multi-instance learning (MIL) was used to integrate prediction across multiple slices to improve the performance of the model. To improve predictive performance, a nomogram combining deep learning features and clinicopathologic parameters was constructed. Model performance was assessed using Harrell's C-index and time-dependent ROC curves.

resultsThe training set included 268 patients, 115 patients in the validation set and 75 patients in the external test set. For OS, the MIL model achieved a C-index of 0.757 in the training cohort, 0.754 in the validation cohort, and 0.741 in the test cohort, compared to 0.666, 0.772, and 0.756 for the clinical model, respectively. The combined model, which integrates MIL features with clinical features, further improved predictive performance, with C-index values for OS at 0.819, 0.822 and 0.759 and for DFS at 0.768, 0.750 and 0.721 across the training, validation and external test cohorts.

conclusionsBy leveraging the complementary strengths of clinical and deep learning approaches, the combined model enhances predictive robustness, enabling more accurate and personalized pretreatment risk assessment in rectal cancer. KEY POINTS: Question Rectal cancer management requires more precise prognostic models to optimize treatment strategies and improve clinical decision-making for individual patients. Findings The combined model leverages synergistic effects between clinical and deep learning features, achieving enhanced prognostic performance and enabling more personalized pretreatment risk stratification. Critical relevance This study demonstrates that MIL extracts deep learning features complementary to clinical knowledge. The combined model leverages this synergy, providing clinicians with a more powerful tool for personalized prognostic assessment.