AI spots hidden danger zones in breast tumors that predict relapse
Researchers used artificial intelligence to map aggressive cell clusters within tumors, finding that where danger signals are located—not just their presence—predicts which patients will relapse. The discovery could reshape how pathologists assess breast cancer risk and guide treatment decisions for thousands of patients annually.
Originaltitel: Abstract PS3-06-04: Spatial representation of deep-learning markers show additional prognostic value in breast cancer patients
Abstract Introduction: Intra-tumour heterogeneity has been hypothesised to increase risk of recurrence in breast cancer patients but has not been studied systematically. Deep-learning enables systematic extraction of prognostic information from H&E histopathology whole slide images (WSI), however, local tile level features are typically aggregated without spatial context. Research objective: To investigate the spatial distribution of an AI-based prognostic marker (DeepGrade) within tumour areas to identify patterns of spatial heterogeneity . Materials and Methods: 3325 WSIs from resected breast tumours of patients diagnosed in two hospitals in Sweden were used to predict DeepGrade status on a tile-level. Tumour regions were segmented and divided spatially into a tumour front and a tumour centre. For each tumour region, the total number of tiles, and the proportion of high-risk tiles were calculated. Clusters of high-risk tiles (regions of 25 neighboring tiles) were defined in the tumour front as a representation for tumour aggressiveness. Analyses were performed by tumour subtypes (Luminal A, Luminal B, Her2-enriched and Basal-like). Cox Proportional Hazards analyses was used to evaluate prognostic performance with progression free survial as primary endpoint. Results: The proportion of DeepGrade-high tiles in the centre area were higher in Her2+ and basal-like patients than luminal patients, (49.6% vs 12.8% had > 80% of centre tiles DeepGrade-high). In the subgroup of luminal patients (2268 patients), those who had a cluster of DeepGrade-high tiles in the tumour front area had higher recurrence propability with a multivariate hazard ratio of 1.97 (CI: 1.19-3.25; p-value=0.008); and within the subgroup of luminal patients with DeepGrade-low status (1356 patients), having a cluster in the tumour front had a univariate hazard ratio of 3.20, and a multivariate hazard ratio of 1.98 (CI: 1.13-3.49, p-value=0.017) when controlling for age, tumour size, lymph node status, and grade. Conclusions: The presence of at least one cluster of high-risk tiles within the tumour front was found to be an independent prognostic factor. More generally, the spatial distribution of high-risk tumour areas in a histopathology slides can have prognostic implications and should be characterised with greater detail in the future. Citation Format: C. Boissin, J. Hartman, M. Rantalainen. Spatial representation of deep-learning markers show additional prognostic value in breast cancer patients [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-06-04.