Evaluation of a Mammography-based Deep Learning Model for Breast Cancer Risk Prediction in a Triennial Screening...

Area of Science:

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

Background:

• Deep learning algorithms demonstrate superior performance in retrospective breast cancer risk assessment compared to traditional methods.
• However, the efficacy of these algorithms in predicting interval cancers (ICs) within triennial screening programs remains under-evaluated.

Purpose of the Study:

• To assess the predictive accuracy of a deep learning algorithm, Mirai, in identifying women who develop interval cancers (ICs) within a 3-year period.
• To evaluate Mirai's performance across different age groups and breast densities within the UK's triennial mammography screening program.

Main Methods:

• A retrospective analysis of 134,217 digital screening mammograms from women aged 50-70 years was conducted.
• The Mirai algorithm generated 3-year risk scores for interval cancers (ICs) based on negative screening mammograms.
• Performance was evaluated using area under the receiver operating characteristic curve (AUC) and C-indexes, comparing predictions at 1, 2, and 3-year intervals.

Main Results:

• Mirai demonstrated consistent predictive performance for ICs across 1, 2, and 3-year intervals, with overall AUCs ranging from 0.67 to 0.72.
• The algorithm's predictive ability showed no significant differences across various age quartiles and Breast Imaging Reporting and Data System (BI-RADS) breast density categories.
• The top 20% of women identified by the 3-year risk scores accounted for 42.4% of all interval cancers detected.

Conclusions:

• The Mirai deep learning algorithm effectively predicts interval cancers (ICs) in women undergoing triennial mammography screening.
• Mirai's risk stratification capabilities can potentially identify women who may benefit from more frequent screening or additional imaging for earlier breast cancer detection.