Sensitivity of a deep-learning-based breast cancer risk prediction model

Area of Science:

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

Background:

• Deep learning models for breast cancer risk (BCR) prediction are increasingly used.
• These models may be sensitive to variations in mammogram acquisition.
• Understanding this sensitivity is crucial for reliable clinical implementation.

Purpose of the Study:

• To investigate the sensitivity of a state-of-the-art BCR prediction model to realistic mammogram image alterations.
• To assess the impact of these alterations on individual BCR predictions and overall model performance.

Main Methods:

• Utilized 5076 mammograms from Slovenian and Belgian screening programs.
• Applied various simulated image alterations (e.g., breast swapping, cropping, rotation, pectoral muscle removal) to the MIRAI model for BCR estimation.
• Evaluated prediction bias, limits of agreement (LOA), and discrimination performance (AUC) using Bland-Altman plots and AUC analysis.

Main Results:

• Breast swapping and inframammary fold alterations had minimal impact.
• Translation, rotation, cropping, and registration caused LOAs up to ±0.1.
• Complete pectoral muscle removal led to substantial prediction bias and wider LOAs.
• No alterations affected the overall discrimination performance (AUC).

Conclusions:

• Mammogram image alterations can cause significant individual variations in predicted breast cancer risk.
• Despite stable overall discrimination, these changes may affect clinical decision-making.
• Further research is needed to ensure the robustness of deep learning BCR models in real-world clinical settings.