Predicting MammaPrint Recurrence Risk from Breast Cancer Pathological Images Using a Weakly Supervised Transformer

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Genomic medicine

Background:

• Recurrence in breast cancer (BC) significantly impacts mortality.
• The MammaPrint (MP) genomic assay assesses recurrence risk and chemotherapy benefit for early-stage HR+/HER2- BC.
• MP has limitations including high cost and inability to analyze tumor morphology.

Purpose of the Study:

• To develop a weakly supervised agent-attention transformer model (CPMP) for predicting MP recurrence risk using histopathological slides.
• To explore spatial and morphological patterns associated with MP risk groups.
• To evaluate the prognostic capability of CPMP in an external cohort.

Main Methods:

• Establishment of a breast cancer MammaPrint cohort.
• Development of CPMP, a weakly supervised agent-attention transformer model.
• Prediction of MP risk groups from annotation-free histopathological slides.
• Spatial and morphological analyses using CPMP.
• Prognostic evaluation in an external cohort.

Main Results:

• CPMP achieved an AUROC of 0.824 ± 0.03 in predicting MP risk groups.
• The model revealed tumor spatial localization and distinct intercellular interaction patterns for different MP risk groups.
• CPMP characterized tumor morphology diversity, identifying unique phenotypes associated with MP risk.
• Prognostic evaluation showed significant stratification of distant metastasis risk (HR: 3.14, p-value = 0.0014).

Conclusions:

• CPMP effectively predicts MammaPrint recurrence risk from histopathological slides.
• The model provides novel insights into spatial and morphological tumor characteristics related to recurrence risk.
• CPMP demonstrates significant prognostic value, offering a cost-effective supplement to genomic risk assessment in early-stage breast cancer.