Deep learning for predicting future lesion emergence in high-risk breast MRI screening: a feasibility study

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• International guidelines recommend contrast-enhanced magnetic resonance imaging (CE-MRI) for high-risk breast cancer (BC) screening.
• The study investigates the use of deep learning for anomaly detection in negative BC screening exams.

Purpose of the Study:

• To assess the applicability of deep learning-based anomaly detection in identifying subtle changes in CE-MRI scans.
• To determine if these anomalies are associated with future lesion emergence in high-risk women.

Main Methods:

• A generative adversarial network was trained on dynamic CE-MRI data from 33 high-risk women without BC.
• An anomaly score was calculated based on deviations from normal breast tissue variability.
• Associations with future lesion emergence were analyzed using ROC curves and logistic regression.

Main Results:

• Local anomaly scores on image patches predicted future lesion emergence with an area under the ROC curve of 0.804.
• An exam-level anomaly score was significantly associated with later lesion emergence (p=0.045).

Conclusions:

• Breast cancer lesions are preceded by detectable anomalous appearance changes in CE-MRI in high-risk women.
• These early image signatures can inform personalized BC risk assessment and screening adjustments.