Breast Cancer Molecular Subtype Prediction: A Mammography-Based AI Approach.

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Breast cancer molecular subtypes significantly impact patient prognosis and treatment.
• Current subtype identification relies on invasive, costly, and time-consuming biopsies, which can be limited by errors and tumor heterogeneity.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model for predicting breast cancer molecular subtypes using mammography images.
• To explore the efficacy of different AI classification strategies and data balancing techniques for subtype prediction.

Main Methods:

• Utilized the OPTIMAM imaging database with 1397 mammograms from 660 patients.
• Employed a pretrained ResNet-101 deep learning model for classifying tumors into five subtypes: Luminal A, Luminal B1, Luminal B2, HER2, and Triple Negative.
• Investigated binary and multi-class classification, alongside data augmentation and resampling techniques to handle imbalanced data.

Main Results:

• Binary classification achieved a maximum average accuracy of 79.02% and AUC of 64.69%.
• Multi-class classification yielded an average AUC of 60.62% with oversampling and data augmentation.
• The HER2 vs. non-HER2 binary classification demonstrated the highest performance with 89.79% accuracy and 73.31% AUC.

Conclusions:

• Mammography-based AI shows significant potential for non-invasive breast cancer molecular subtype prediction.
• This AI approach could serve as a valuable alternative to traditional biopsies, facilitating personalized treatment planning.
• Further development could enhance diagnostic accuracy and streamline breast cancer management.