Deep Learning on Histopathological Images to Predict Breast Cancer Recurrence Risk and Chemotherapy Benefit

Area of Science:

• Oncology
• Artificial Intelligence in Medicine
• Pathology

Background:

• Genomic testing is crucial for early breast cancer treatment but faces accessibility issues globally.
• Hormone receptor-positive, HER2-negative (HR+/HER2-) early breast cancer treatment decisions are often guided by genomic scores.
• High costs and logistical barriers limit access to genomic testing for many patients worldwide.

Purpose of the Study:

• To develop an artificial intelligence (AI) model to estimate Oncotype DX 21-gene recurrence scores (RS) using histopathology images and clinicopathologic data.
• To assess the AI model's accuracy in predicting genomic risk and its prognostic value in HR+/HER2- early breast cancer.
• To evaluate the AI model's utility in guiding chemotherapy decisions, especially in resource-limited settings.

Main Methods:

• A multimodal deep learning AI model was developed, pre-trained on a large dataset of histopathological slides.
• The AI model was fine-tuned and validated using data from the TAILORx randomized trial (n=8,284).
• External validation was performed on six independent cohorts (n=5,497 patients) to assess generalization.

Main Results:

• The AI model accurately estimated recurrence scores, achieving an AUC of 0.898 for predicting high genomic risk (RS≥26).
• AI-driven risk stratification demonstrated significant prognostic value for recurrence-free and disease-free survival.
• The model identified chemotherapy benefit in premenopausal high-risk patients and ruled it out in postmenopausal low-risk patients, aligning with clinical trial data and reclassifying ~30% of MINDACT high-risk postmenopausal cases as low-risk.

Conclusions:

• AI applied to standard histopathology images provides a valuable, accessible tool for chemotherapy decision-making in HR+/HER2- early breast cancer.
• This AI approach can help reduce unnecessary chemotherapy and advance precision medicine, particularly in settings with limited access to genomic testing.
• The AI model demonstrates robust generalization and prognostic capability, offering a scalable solution for personalized cancer care.