Benchmarking Distance Functions in Siamese Networks for Current and Prior Mammogram Image Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Mammogram image analysis is crucial for early breast cancer detection.
• Artificial intelligence (AI), specifically Siamese networks, shows promise in comparing current and prior mammograms.
• Selecting an effective distance function is a key challenge for Siamese networks in this application.

Purpose of the Study:

• To explore the impact of non-linear and correlation-sensitive distance functions in Siamese networks for mammogram analysis.
• To benchmark various distance functions and introduce a novel combination for improved performance.
• To enhance the diagnostic accuracy and generalization capabilities of AI in mammography.

Main Methods:

• Implemented and evaluated several distance functions: Euclidean, Manhattan, Mahalanobis, Radial Basis Function (RBF), and cosine.
• Introduced and tested a novel distance function: RBF combined with Matern Covariance.
• Benchmarked performance using metrics such as accuracy, sensitivity, precision, specificity, F1 score, and AUC on paired mammogram images.

Main Results:

• The RBF with Matern Covariance distance function consistently outperformed traditional functions.
• The ResNet50 model with the proposed distance function achieved high performance metrics (e.g., accuracy 0.938, AUC 0.940).
• The approach demonstrated robustness and generalizability across 30 cross-validation samples.

Conclusions:

• Non-linear and correlation-based distance functions are vital for effective Siamese network performance in mammogram analysis.
• The RBF with Matern Covariance offers a superior method for capturing subtle differences in correlated mammogram images.
• This research advances AI-driven mammography, potentially leading to more accurate and reliable diagnostic tools.