Multiscale characterization of early morphologic radiation response in rectal cancer via digital pathology

With effective neoadjuvant chemotherapy and radiotherapy (RT), some patients with locoregional rectal cancer may avoid surgery. However, patient responses are highly heterogeneous and poorly understood. Although morphological assessment remains the clinical gold standard, it is often dominated by tumor abundance, potentially overlooking complex biological adaptations to treatment. To address this, we sought to gain insights through a multiscale characterization of early histologic changes following RT. We leveraged matched pre- and post-RT histology slides from the INNATE trial dataset, using adjacent tissue as a control, to develop a dual-pipeline framework: (1) C-MorphQuant, for analysis of predefined classical features such as tissue composition, spatial organization, and nuclear morphology based on a finetuned cell classifier and a trained region classifier; and (2) DL-RadScore, a novel deep learning metric that learns de novo morphological features by distinguishing irradiated from non-irradiated tissue. Our C-MorphQuant analysis revealed: (1) compositionally, a clear reduction in tumor and normal epithelium with a commensurate increase in stroma; (2) spatially, the disruption of glands into fragmented small clusters (<30 nuclei); and (3) nuclearly, alterations in size and shape that were evident but patient-specific. Notably, while the overall radiation impact was generally stronger in tumor, C-MorphQuant phenotypes were more heterogeneous than the stereotypical adaptations observed in adjacent normal tissue. The DL-RadScore accurately distinguished unseen pre- and post-radiation samples. We established biological relevance by comparing with tumor epithelial Ki67 proliferation and found strong negative correlations at both the patient (r = -0.59, p = 2.08*10 -2 ) and intra-slide (r = -0.43, p = 2.17*10 -78 ) levels. While DL-RadScore maintained strong concordance with C-MorphQuant composition metrics, it outperformed measures of tumor cellularity in capturing proliferative status. Most intriguingly, our results demonstrated that rather than a gradual accumulation of morphological changes, some tumors appeared to initially "shrug off" early radiation doses, preserving a pre-treatment-like morphology, yet ultimately collapsed by the time of surgery. In summary, this study underscores the value of characterizing treatment-induced phenotypes beyond mere tumor cell abundance and highlights the complex non-linear spatiotemporal dynamics in which early morphological persistence does not preclude long-term cure.