Bilateral Analysis Boosts the Performance of Mammography-based Deep Learning Models in Breast Cancer Risk Prediction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality in women, necessitating improved early detection methods.
• Breast density is a known risk factor, and deep learning models, particularly Convolutional Neural Networks (CNNs), show promise in analyzing mammograms.
• Current risk prediction models often analyze mammograms unilaterally, potentially missing valuable comparative information.

Purpose of the Study:

• To enhance breast cancer risk prediction by utilizing bilateral analysis, comparing mammograms from both breasts of a patient.
• To develop and evaluate a Siamese neural network designed to detect asymmetries indicative of risk.
• To compare the performance of the Siamese model against traditional unilateral CNN models and explore ensemble methods.

Main Methods:

• Development of a Siamese neural network architecture to process and compare bilateral mammograms.
• Training and testing the Siamese model on a dataset of 271 patients.
• Comparative analysis of the Siamese model's performance against a unilateral CNN model using metrics like AUC, sensitivity, specificity, and precision.
• Implementation of ensemble techniques (pre-trained weights, weighted voting) to merge both models.

Main Results:

• The Siamese neural network achieved a higher Area Under the Curve (AUC) of 0.75 compared to the unilateral CNN model's AUC of 0.70 (p = 0.0056).
• The Siamese model demonstrated superior performance across sensitivity (0.68 vs. 0.61), specificity (0.69 vs. 0.66), and precision (0.71 vs. 0.67), with a lower false positive rate (0.31 vs. 0.34).
• Ensemble methods combining the Siamese and CNN models further boosted the AUC to 0.78, indicating synergistic benefits.

Conclusions:

• Bilateral analysis using Siamese neural networks significantly improves breast cancer risk prediction compared to traditional unilateral approaches.
• Leveraging inter-breast asymmetries in mammograms is a valuable strategy for enhancing the accuracy of deep learning-based risk assessment.
• Ensemble techniques integrating bilateral and unilateral models offer a promising avenue for further advancements in breast cancer detection and risk stratification.