Longitudinal Analysis of Changes in Deep Learning Image-based Breast Cancer Risk Scores over Time

Area of Science:

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

Background:

• Artificial intelligence (AI)-based deep learning (DL) models offer individualized 5-year breast cancer risk estimates from screening mammograms.
• While validated for static risk prediction, the temporal behavior of these AI risk scores is not well-established.

Purpose of the Study:

• To investigate the longitudinal changes in image-based AI deep learning (DL) risk scores.
• To determine if risk score trajectories differ between women who develop breast cancer and those who remain cancer-free.

Main Methods:

• Retrospective cohort study of 158,807 screening mammograms from 54,014 women (2009-2019).
• Comparison of 817 women diagnosed with breast cancer within one year against cancer-free controls.
• Validated image-only DL model used to generate continuous 5-year risk scores; linear mixed-effects models analyzed score trajectories.

Main Results:

• In women who developed cancer, median AI risk scores increased from 2.1 (6 years pre-diagnosis) to 6.6 (index exam).
• Cancer-free women exhibited stable scores (1.8-2.2) over time.
• Longitudinal analysis showed significant score increases in the cancer group (slope 1.13/year) versus minimal change in controls (slope 0.09/year).

Conclusions:

• AI-based breast cancer risk scores derived from mammograms demonstrate dynamic evolution over time.
• Diverging score trajectories between cancer and cancer-free groups highlight their potential as dynamic biomarkers.
• These findings support the use of AI risk scores for risk-adaptive screening and prevention strategies.