MR-DELTAnet: A Longitudinal MRI-Transformer Model Predicting Pathological Complete Response and Revealing Immune Microenvironment via scRNA-seq in Locally Advanced Rectal Cancer.

Accurate tumor response assessment to neoadjuvant chemoradiotherapy (NCRT) is crucial for personalized treatment strategies in locally advanced rectal cancer (LARC). However, reliable non-invasive assessment tool remains clinically lacking. To fill this unmet need, MR-DELTAnet, a longitudinal MRI-based Transformer framework that integrates Delta-Efficient Latent-Temporal Attention, is constructed to predict pathological complete response (pCR) to NCRT in locally advanced rectal cancer patients. In a multicenter retrospective cohort of 1,026 LARC patients between July 2012 and July 2023, MR-DELTAnet demonstrated robust discriminative performance across independent datasets, with the area under the curves (AUC) of 0.93 (95% CI 0.90-0.96), 0.88 (95% CI 0.82-0.94) and 0.90 (95% CI 0.79-1.00) and in training (n═633), internal validation (n═212) and external validation (n═181) sets, respectively. Risk-stratification by MR-DELTAnet prediction scores reveals significant survival differences: low-score patients exhibit prolonged disease-free and overall survival versus high-score patients (log-rank p<0.05). Applying the model to an independent single-cell RNA sequencing cohort (n═26) discloses biologically distinct immune microenvironments: high-score tumors are myeloid-rich and immunosuppressive, whereas low-score tumors harbor cytotoxic T-cell-dominant. Clinically, MR-DELTAnet provides an accurate, non-invasive tool for preoperative identification of pCR likelihood and biological phenotype, thereby potentially informing individualized treatment strategies for LARC management.