Explainable Multiparametric MRI Radiomics Reveals Prognostically Relevant Intratumoral Heterogeneity in Rectal Cancer.

Rationale

AND

Objectives

Clinically relevant intratumoral heterogeneity (ITH) in rectal cancer remains difficult to quantify noninvasively using conventional visual assessment of pretreatment magnetic resonance imaging (MRI). We developed an explainable multiparametric MRI radiomics signature for ITH-based risk stratification.

Materials And Methods

This retrospective multicenter study included 794 patients, with 600 in the training cohort and 194 in two external validation cohorts. Radiomics features from T2-weighted imaging, diffusion-weighted imaging, and apparent diffusion coefficient maps were integrated using similarity network fusion to identify MRI heterogeneity states associated with disease-free survival (DFS). An extreme gradient boosting model with Shapley additive explanations (SHAP) was then used to derive an 18-feature texture-based ITH score, which was dichotomized into high and low SHAP heterogeneity.

Results

High SHAP heterogeneity was associated with worse DFS in the training cohort (P < 0.001), external validation cohort 1 (P = 0.040), and external validation cohort 2 (P = 0.010). It remained independently associated with inferior DFS in the training cohort (hazard ratio [HR], 1.614; 95% confidence interval [CI], 1.026-2.536; P = 0.038) and external validation cohort 1 (HR, 2.555; 95% CI, 1.042-6.261; P = 0.040). In patients receiving neoadjuvant therapy, high SHAP heterogeneity was independently associated with lymphovascular invasion (odds ratio, 2.523; 95% CI, 1.103-5.886; P = 0.029). A nomogram integrating SHAP heterogeneity with carcinoembryonic antigen and MRI-T stage improved 3-year DFS prediction.

Conclusion

High SHAP heterogeneity identified an aggressive MRI phenotype linked to poor DFS and residual invasive pathology, supporting this explainable multiparametric MRI radiomics approach for pretreatment risk stratification in rectal cancer.