Un modelo no invasivo de IA basado en características de RM multidimensionales para predecir estructuras linfoides...

Tertiary lymphoid structures (TLSs) in rectal cancer (RC) are closely associated with immunotherapy response and patient prognosis, yet their assessment currently relies on invasive biopsy. This study aimed to investigate whether TLSs can be accurately and noninvasively predicted using multi-dimensional MRI features, and to evaluate the model's utility in predicting immunotherapy response and prognosis.

A total of 606 RC patients from four cohorts were included. A development cohort was used for model construction, while both a validation cohort and a prospective cohort were used to assess its generalizability. An additional immunotherapy cohort was utilized to evaluate the model's ability in predicting immune responses. The proposed multi-dimensional features comprised: (1) radiomic features extracted from ROI; (2) dimensionality-reduced features derived using principal component analysis (PCA) and singular value decomposition (SVD); and (3) tumor heterogeneity features extracted via habitat analysis. XGBoost was employed to construct the TLSs classification model (positive vs. negative). Shapley Additive exPlanations (SHAP) analysis was used to interpret the contributions to model decisions, and the model's performance was further tested in predicting immunotherapy response and survival outcomes.

The TLSs model demonstrated strong discriminatory performance, with AUROC of 0.88 in the internal validation set, 0.81 in external test set 1, and 0.84 in external test set 2. Features such as "subregion3_firstorder_MeanAbsoluteDeviation" had the greatest impact on model decisions. Furthermore, the TLSs score derived from the model showed promising predictive value for pathological complete response to immunotherapy, with an AUROC of 0.74. Kaplan-Meier analysis revealed that the high TLSs score group had significantly better disease-free survival (DFS) compared to the low-score group.

The multi-dimensional MRI-based TLSs model shows robust performance in predicting TLSs status, immunotherapy response, and prognosis in RC, providing a novel tool for guiding personalized immunotherapy and prognostic assessment.