<p>Supplementary methods, figures and tables.</p>

Fig C. Construction of the segmentation training dataset and five-fold cross-validation workflow for the AI-mrEMVI model. Fig D. The performance of the AI-mrEMVI model across different Dice scores. Fig E. Representative comparisons of AI-mrEMVI model predictions and radiologist annotations in typical, ambiguous, and misclassified cases.

Fig F. Site-stratified confusion matrices and performance metrics of the AI-mrEMVI model in external test cohort 2. Fig G. The prognostic value of AI-mrEMVI status. Fig H. Bootstrap-derived distributions of C-index for DFS and OS across three models.

Fig I. Kaplan-Meier survival curves for discordant cases between AI-mrEMVI and radiologist ground truth assessment. Fig J. The prognostic value of AI-mrEMVI lesion volume in external test cohort 2. Fig K. Correlation and agreement between AI-derived and manually annotated mrEMVI lesion volumes in true-positive cases.

Fig L. Kaplan-Meier curves for DFS according to AI-mrEMVI status in different mrT and mrN staging subgroups. Fig M. Kaplan-Meier curves for DFS according to AI-mrEMVI status in different tumor location subgroups. Fig N. Kaplan-Meier curves for DFS according to AI-mrEMVI status in different subgroups of MR T2WI images with varying slice thicknesses and spacing between slices.

Fig O. Kaplan-Meier curves for DFS according to AI-mrEMVI status in different subgroups of MR T2WI images with varying pixel spacing. Table A. MRI acquisition parameters across the participating centers. Table B. Segmentation performance comparison across deep learning architectures in internal five-fold cross-validation.

Table C. Classification performance comparison across deep learning architectures on the combined external test cohorts. Table D. Delta C-index and 95% confidence intervals for model comparisons in disease-free survival (DFS) and overall survival (OS). Table E. Univariable and multivariable analyses of OS.

Table F. Schoenfeld residual test results for proportional hazards assumption in multivariable Cox regression models of DFS and OS in two external test cohorts.</p> <p>(PDF)</p>