Domain Generalization Mitigates Scanner-Induced Domain Shift in Medical Imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models in medical imaging face performance degradation due to domain shift from varied acquisition hardware and protocols.
• This limits the clinical deployment of AI tools for tasks like cancer grading and density assessment.

Purpose of the Study:

• To comprehensively evaluate domain generalization (DG) techniques for mitigating performance degradation in medical image analysis.
• To assess the effectiveness of DG methods on distinct clinical tasks using multi-institutional datasets.

Main Methods:

• Evaluation of six DG algorithms against a baseline using a leave-one-domain-out protocol.
• Testing on two tasks: grading prostate cancer aggressiveness from MRI (ProstateNet dataset) and assessing breast density from mammograms (DMIST dataset).

Main Results:

• DG methods, especially those with explicit regularization, improved out-of-domain generalization.
• The FISH algorithm achieved the highest out-of-domain AUROC (0.678) on the ProstateNet dataset, a significant improvement over the baseline (0.613).
• Similar trends were observed on the DMIST dataset, though DG did not fully close the gap with in-domain performance.

Conclusions:

• Domain generalization strategies are necessary for developing clinically deployable AI models in medical imaging.
• While DG methods enhance robustness to domain shift, further research is needed to fully bridge the gap with in-domain performance.