Deep Learning Identifies BMI-Associated Morphologic Features in Breast Cancer Tissue

Multiple epidemiological studies have established obesity as an adverse factor in breast cancer incidence, progression, and clinical outcomes. However, the biological mechanisms underlying these associations remain incompletely understood. In peritumoral adipose tissue, obesity is associated with chronic inflammation, macrophage recruitment, and broader alterations in the local microenvironment. Whether these obesity-associated changes are accompanied by distinct morphologic features within the tumor and non-adipose stromal compartments remains unclear. In this study, we investigated whether an AI-based pathology foundation model could support the development of machine learning algorithms capable of identifying obesity-associated morphologic patterns in histopathology images, with a particular focus on tumor and stromal regions. In a cohort of 195 patients with breast cancer, machine learning models using vector embeddings from non-adipose stromal regions classified high BMI ([≥]30 kg/m2) versus low BMI (<25 kg/m2) with an average AUC of 0.78 (95% CI: 0.73-0.83), consistently outperforming models based on tumor-only or combined tumor-stroma regions. These obesity-associated stromal signatures were observed across subgroups defined by hormone-receptor status and age, indicating that they are not restricted to specific clinical populations. Complementary embedding-based dissimilarity analysis further showed that higher BMI was associated with reduced inter-patient heterogeneity in both tumor and stromal tissues. Overall, these findings suggest that obesity is associated with distinct morphologic alterations in both tumor and non-adipose stromal tissues. Notably, the stromal compartment emerged as a particularly informative indicator of obesity-associated morphology and demonstrated reduced inter-patient heterogeneity among individuals with high BMI.