Domain generalization for mammographic image analysis with contrastive learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Deep learning models require diverse datasets for effective mammography analysis.
• Collecting diverse mammography data from various vendors is practically challenging.
• Existing methods struggle with style generalization in deep learning for medical imaging.

Purpose of the Study:

• To develop a novel contrastive learning method (MSVCL+) for improved style generalizability in deep learning models for mammography.
• To enhance the robustness of feature embeddings against variations in mammogram data styles.
• To improve the performance of computer-aided diagnosis tasks using a generalizable pretrained model.

Main Methods:

• Developed MSVCL+, a multi-style, multi-view unsupervised self-learning scheme for pretraining.
• Utilized contrastive learning to achieve robust feature embedding against style diversity.
• Fine-tuned the pretrained network for downstream mammography tasks: mass detection, matching, BI-RADS rating, and breast density classification.

Main Results:

• The MSVCL+ method demonstrated significant improvements in four mammographic image analysis tasks.
• The approach effectively generalized to unseen domains (data from different vendor styles).
• Outperformed several state-of-the-art domain generalization methods on public datasets.

Conclusions:

• MSVCL+ provides a robust solution for training generalizable deep learning models in mammography.
• The method addresses the challenge of data diversity by enhancing style generalizability.
• This approach holds promise for improving the accuracy and reliability of computer-aided diagnosis in mammography.