Temporal Machine Learning Analysis of Prior Mammograms for Breast Cancer Risk Prediction.

Area of Science:

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

Background:

• Identifying women at risk for sporadic breast cancer is a significant clinical challenge.
• Traditional risk assessment methods may not fully capture the dynamic nature of breast cancer development.

Purpose of the Study:

• To investigate the efficacy of temporal analysis of annual screening mammograms using a long short-term memory (LSTM) network for identifying women at risk of future breast cancer.
• To evaluate the performance of deep-learning and radiomic features in predicting lesion malignancy.

Main Methods:

• A case-control study retrospectively collected sequences of antecedent mammograms from women with biopsy-confirmed abnormalities.
• Radiomic and deep-learning features were extracted from regions of interest in antecedent images.
• LSTM recurrent networks were employed to classify future lesions as malignant or benign using single and multiple time-points.

Main Results:

• Classifiers incorporating multiple time-points with LSTM, using either deep-learning or radiomic features, performed significantly better than chance (AUCs ranging from 0.63 to 0.65).
• Classifiers using only a single time-point did not show improved performance compared to chance (AUCs around 0.52-0.54).
• Similar classification performance was observed when analyzing features from the affected versus the contralateral breast, suggesting a potential field effect.

Conclusions:

• Incorporating temporal information into radiomic analyses via LSTM networks can improve the classification performance for predicting future breast cancer lesions.
• The presence of a potential field effect in antecedent imaging, observable in both breasts, indicates that evaluating either breast may inform future breast cancer risk.