Spatiotemporal Mammography-based Deep Learning Model for Improved Breast Cancer Risk Prediction

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Breast cancer is a leading cause of death in women, necessitating improved early detection and risk assessment.
• Traditional risk models show population inconsistencies, limiting their clinical utility.
• Deep learning, specifically Convolutional Neural Networks (CNNs), has shown promise in breast cancer risk prediction from mammograms.

Purpose of the Study:

• To enhance breast cancer risk prediction by incorporating spatiotemporal information from multiple screening mammograms.
• To evaluate a Siamese neural network model for spatiotemporal risk prediction against traditional CNN models.

Main Methods:

• A Siamese neural network was developed to analyze spatiotemporal patterns in screening mammograms.
• The Siamese network's performance was compared to CNNs trained on individual time points (T1 and T2).
• The models were tested on a dataset of 191 cases, including 61 diagnosed with breast cancer.

Main Results:

• The Siamese network achieved a superior Area Under the Curve (AUC) of 0.81, compared to 0.75 (T1) and 0.77 (T2) for CNNs.
• The Siamese model demonstrated higher accuracy (0.78) and F1-score (0.61) than CNNs (Accuracy: 0.76 (T1), 0.76 (T2); F1-score: 0.54 (T1), 0.59 (T2)).

Conclusions:

• Spatiotemporal analysis of mammograms using a Siamese network significantly improves breast cancer risk prediction.
• This approach offers a more reliable tool for personalized risk assessment compared to models using single time points.