An Analysis-First Framework for Region-Level Heterogeneity Characterization in Breast Cancer Whole-Slide Images
Histopathological whole-slide images (WSIs) contain rich spatial information reflecting tissue organization and heterogeneity, yet many computational approaches reduce these images to patch-level predictions or coarse aggregates, limiting tissue-scale interpretability. This study presents an analysis-first, region-aware framework for characterizing tissue heterogeneity in breast cancer WSIs. Diagnostic slides from The Cancer Genome Atlas Breast Invasive Carcinoma (TCGA-BRCA) cohort were partitioned into spatially localized tissue regions, and region-level representations were extracted using a frozen ResNet-50 encoder. Without relying on supervised learning, a region-aware heterogeneity characterization framework was developed using four representation-space measures capturing regional feature dispersion, inter-region dissimilarity, embedding-norm concentration, and clustering-derived rarity. Evaluation on 173 WSIs revealed non-degenerate metric distributions and structured correlations, indicating that tissue organization cannot be fully summarized by a single representation-space measure. Quantile-based thresholding identified recurring organizational phenotypes termed concentrated low dispersion, diffuse heterogeneous, and rare pattern rich. Representative cases illustrated visually distinct appearances associated with the metric-based phenotype definitions. Robustness analyses demonstrated stable phenotype assignments under moderate threshold perturbations. In addition, region-size ablation experiments showed that dispersion and embedding-norm concentration measures remained highly consistent across spatial scales, whereas clustering-based rarity measures were more sensitive to region granularity. A pilot encoder-sensitivity analysis further showed moderate-to-strong preservation of metric rankings between ResNet-50 and pathology-specific UNI representations. The proposed framework explicitly separates heterogeneity characterization from predictive modeling and provides an interpretable approach for studying tissue organization in WSIs. These results provide a basis for future studies on tissue heterogeneity using computational pathology.
Authors
- Mohammad Mehedi Hassan (ORCID: https://orcid.org/0000-0002-3479-3606)
Institutions
- Kyoto College of Graduate Studies for Informatics (JP)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11763682/v1
- Primary Topic
- AI in cancer detection
- Type
- preprint