Longitudinal interpretability of deep learning based breast cancer risk prediction

Area of Science:

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

Background:

• Deep-learning models achieve state-of-the-art breast cancer risk (BCR) prediction.
• Understanding the mechanisms of BCR prediction is crucial for clinical trust and application.
• Key questions remain on whether these models detect morphologic changes indicative of cancer.

Purpose of the Study:

• To determine the timeframe during which oncogenic processes provide sufficient signal for BCR models to detect changes.
• To enhance the interpretability of deep-learning-based BCR prediction models.
• To provide clinicians with new perspectives on model behavior.

Main Methods:

• Utilized MIRAI, a BCR risk prediction model, on 1210 screening mammograms from patients screened pre-diagnosis and 2400 from patients with long-term follow-up.
• Defined attribution heterogeneity as the relative difference in model attributions between breasts using eight interpretability techniques.
• Quantified model reliance on the cancer-affected breast side using AUC and compared attribution heterogeneity between cancer patients and healthy individuals via Mann-Whitney U test.

Main Results:

• Model reliance on the cancer side was highest for 0-1 years-to-cancer (AUC=0.85-0.95), decreased for 1-3 years (AUC=0.64-0.71), and remained above random for 3-5 years (AUC=0.51-0.58).
• All eight attribution methods showed significantly larger median absolute attribution heterogeneity in cancer patients (p<0.01).
• Longitudinal trends in attribution heterogeneity were consistent across tested methods.

Conclusions:

• Long-term BCR predictions (>3 years) likely rely on typical breast characteristics like density.
• Mid-term predictions (1-3 years) suggest the model detects early tumor development signs.
• Short-term predictions (≤1 year) indicate the BCR model functions as a breast cancer detection tool.