Artificial Intelligence for Image-Based Breast Cancer Risk Prediction Using Attention.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Accurate breast cancer risk prediction is crucial for personalized prevention and early detection.
• While genetic data and breast density improve risk models, mammogram image features are underutilized.
• Deep learning can extract complex information from mammograms, but data scarcity and computational demands pose challenges.

Purpose of the Study:

• To develop an attention-based Multiple Instance Learning (MIL) model for accurate, short-term breast cancer risk prediction using full-resolution mammograms.
• To incorporate image-based features from mammograms into risk assessment models.
• To address data scarcity by mixing current screen-detected cancers with prior mammograms during model development.

Main Methods:

• An attention-based Multiple Instance Learning (MIL) model was developed to analyze mammograms.
• The model was trained using a dataset that included prior mammograms from women who later developed screen-detected or interval cancers.
• Performance was evaluated using the Area Under the Curve (AUC) metric.

Main Results:

• The proposed MAI-risk model achieved an AUC of 0.747 [0.711, 0.783] for predicting short-term breast cancer risk.
• This performance surpassed established models like IBIS (AUC 0.594 [0.557, 0.633]) and VAS (AUC 0.649 [0.614, 0.683]).
• The model effectively identified women who developed cancer between 5 and 55 months after their screening mammogram.

Conclusions:

• Deep learning models, specifically attention-based MIL, can accurately predict short-term breast cancer risk from mammograms.
• Incorporating detailed image features from mammograms significantly improves risk prediction accuracy.
• This approach holds promise for enhancing personalized breast cancer screening and prevention strategies.