Multi-scale histology reveals localized ageing signatures in normal human colon

Abstract

Ageing is accompanied by heterogeneous changes in tissue structure, cellular composition, and microenvironments, yet how these changes are reflected in routine human histology remains poorly understood. Here, we introduce a histology-first, multi-scale digital pathology framework to study human colon ageing using hematoxylin and eosin (H&E) whole-slide images from clinically normal colon biopsies. The framework combines pathology foundation-model patch embeddings, attention-based multiple-instance learning, cell-level morphometric profiling, and prototype-based tissue compartmentalization. We show that chronological age can be predicted from normal colon tissue using mesoscopic representations derived from pathology foundation-model patch embeddings and attention-based multiple-instance learning. Incorporating cell-type-level morphometric features further improves prediction and refines the spatial allocation of model attention. By integrating attention maps with prototype-defined tissue compartments, we identify localized age-associated differences within stromal, epithelial, and lymphoid mucosal niches. Finally, we show that prediction discordant cases exhibit cellular profiles shifted toward the model-assigned age group, suggesting that some apparent model errors may reflect tissue-level age discordance beyond chronological age. Together, these findings show that routine H&E histology captures spatially localized signatures of normal human colon ageing and provides a scalable framework for linking tissue morphology with biological ageing state.