Deep learning as a predictor for MammaPrint risk group from pathology slides in Chinese early-stage breast cancer.

Abstract

e13670 Background: MammaPrint (MP) test has been used to evaluate the prognosis and chemotherapy benefit of hormonal receptor positive human epidermal receptor 2 negative (HR+/HER2-) breast cancer (BC). However, its application is constrained by specific clinical scenarios and the high cost of molecular testing. Given that pathological whole slide images (WSIs) provide cellular spatial and morphological information that may reflect phenotypic characteristics of molecular targets, the aim of this study was to identify morphological patterns associated with molecular MP results and to establish an AI model to predict the MP risk group of Chinese HR+/HER2- BC patients from WSIs. Methods: We collected a HR+/HER2- BC cohort from Tianjin Medical University Cancer Hospital (TJMUCH), comprising clinicopathologic features, digital WSIs, and MP results of 477 patients. A multi-instance weakly-supervised transformer model capable of aggregating morphological and spatial features was further developed to predict MP recurrence risk from annotation-free WSIs of BC. Specifically, our AI model was further leveraged for spatial and morphological analysis to explore histological patterns related to MP risk groups. Results: Our AI model shows promising robustness in clinical cohorts, which achieved an average AUC value of 0.825 (5-time 5-fold cross validation) in predicting MP risk group (Low vs. High) from WSIs. Key morphological features, such as dense tumor cellularity, tumor infiltrating lymphocytes, were identified as predictors of MP risk group. The distinct cellular spatial compositions in representative regions elucidate the spatial-morphological patterns associated with prognostic molecular MP testing. Conclusions: Overall, our work established a cost-effective AI model for HR+/HER2- BC patients to assess recurrence risk and guide treatment decisions. Future research will focus on external validation using larger cohorts, integrating this approach as a flexible supplement in routine clinical workflows.