Quantitative Assessment Tumor Regression and Spatial Immune Patterns after Ncrt In Rectal Cancer Using a Deep Learning Framework
Abstract
Purpose
Accurate assessment of locally advanced rectal cancer (LARC) following neoadjuvant chemoradiotherapy (nCRT) is critical for treatment stratification. However, manual Tumor Regression Grade (TRG) suffers from inter-observer variability and subjective bias. This study proposes an objective framework to quantify nCRT response through AI-derived tumor bed ratio (TBR) and spatial tumor-infiltrating lymphocyte (TIL) mapping.
Methods
In this retrospective study, a cohort of 482 LARC patients treated with nCRT followed by surgery was analyzed using a tile-based Vision Transformer (ViT) architecture. The model was trained to segment whole-slide images (WSIs) into five distinct tissue categories, including tumor cells, tumor bed, lymphocytes, normal epithelium and other irrelevant tissues. We quantified the nCRT-induced treatment response by calculating the AI-derived TBR, defined as the volumetric proportion of residual tumor within the fibrotic tumor bed. Furthermore, the nCRT-altered immune microenvironment was characterized using a morphological dilation algorithm that extracted tumor-infiltrating lymphocyte (TIL) densities at incremental spatial distances, ranging from 0.28 mm to 3.32 mm from the tumor margin. An integrated "immuno-TRG" (iTRG) model was subsequently established by combining the TBR with the most prognostic spatial TIL feature.
Results
The AI classifier achieved exceptional performance (AUC: 0.984–1.000). The AI-derived TBR outperformed manual TRG in prognostic accuracy, showing higher C-indices in both training (0.632 vs. 0.587) and validation cohorts (0.630 vs. 0.574). Notably, confusion matrix analysis revealed that AI detected microscopic residual disease in 60% of cases manually classified as pathological complete response (TRG0). Furthermore, 63% of subjective TRG2 cases were reclassified into more granular risk groups, resolving the inherent ambiguity of intermediate manual grading.
Conclusion
In conclusion, integrating AI-derived TBR and spatial TIL metrics provides a reproducible, quantitative methodology for evaluating nCRT response. By converting visual patterns into discrete physical variables, this framework provided a standardized basis for personalizing post-nCRT clinical management and radiation de-escalation strategies.