1076 - Spatiotemporal AI Morphometry for Early Recurrence Assessment in Rectal Cancer After Neoadjuvant Chemoradiotherapy: A Multi-Model Ablation Study
Presenter(s)
S. Chen1, M. Zhang2, J. Wang3, Y. Lin4, Z. Zhang4, L. Wang5, Z. Dong3, Y. Lu5, J. Zhou6, A. Dekker7, P. Kalendralis7, L. Wee8, and Z. Zhang3; 1Department of Radiation Oncology, Fudan University Shanghai Medical College, Shanghai, China, 2Department of Pathology, Fudan University Shanghai Cancer Center, Shanghai, China, 3Department of Radiation Oncology, Fudan University Shanghai Cancer Center; Department of Oncology, Shanghai Medical College, Fudan University; Shanghai Clinical Research Center for Radiation Oncology; Shanghai Key Laboratory of Radiation Oncology, Shanghai, China, 4Department of Radiation Oncology, Fudan University Shanghai Cancer Center, Shanghai, China, 5Fudan University Shanghai Cancer Center, Shanghai, China, 6Maastricht University, Maastricht, Netherlands, 7Department of Radiation Oncology (Maastro), GROW School for Oncology and reproduction, Maastricht University Medical Centre, Maastricht, Netherlands, 8Clinical Data Science, Faculty of Health, Medicine and Life Sciences, Maastricht University, Maastricht, Netherlands
Purpose/Objective(s): Conventional Pathological Tumor Regression Grade (TRG) is limited by interobserver variability and fails to capture the spatial immune-stromal landscape. This study developed an AI-based system to quantify residual tumor morphometry and tumor microenvironment (TME) metrics to improve Disease-Free Survival (DFS) prediction.
Materials/Methods: Whole-slide images from 845 patients with locally advanced rectal cancer post-neoadjuvant chemoradiotherapy (nCRT) were analyzed. A physics-based AI pipeline extracted 16 features across tumor-intrinsic and TME-extrinsic domains. Five independent models were benchmarked: Model A (Clinical TRG via Cox), Models B-D (XGBoost for Tumor-only, TME-only, and Integrated AI). Performance was evaluated via time-dependent C-indices across Early (0-12m) and Overall windows. Feature contributions were decoded using SHAP (SHapley Additive exPlanations).
Results: Univariate analysis identified several robust prognosticators with significant training set p-values. In the Early window (0-12m), gland formation differentiation entropy (H_diff) (P=0.032) and differentiation variation (Diff_Std) (P=0.019) achieved C-indices of 0.603 and 0.611 in the training cohort, respectively. The novel metric Env-Diff Heterogeneity (D_std_env) demonstrated superior early-warning performance (P=0.042), with C-indices of 0.614 and 0.692 in the training and validation cohorts, respectively. In the ablation study, the Integrated AI Model (Model D) achieved the highest overall accuracy, with C-indices of 0.766 and 0.699 in the training and validation cohorts, respectively, significantly outperforming the Clinical TRG baseline (Model A: validation C-index 0.601). SHAP analysis confirmed that spatial metrics, specifically tumor-stroma interaction (TIS, importance: 0.128) and immune-cell ratio (ICR_Max, 0.112), were the primary drivers of the integrated model.
Conclusion: AI-assisted landscape morphometry enables quantitative assessment of both residual tumor and immune response. Integrating AI-derived tumor and TME features significantly outperforms traditional TRG in predicting DFS. This spatiotemporal profiling, particularly through early-sensitive metrics like D_std_env, offers a promising tool for refining postoperative risk stratification and guiding personalized adjuvant therapy.