ROBUST WHITE MATTER HYPERINTENSITY SEGMENTATION ON UNSEEN DOMAIN

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Machine learning models often assume identical training and testing data distributions.
• Medical imaging datasets from multiple sites or scanners violate this assumption due to systematic variability.
• This limits the generalizability of models to new, unseen data distributions.

Purpose of the Study:

• To address the challenge of applying machine learning models to unseen medical imaging data.
• To investigate Domain Generalization (DG) techniques for robust model performance.
• To improve white matter hyperintensity (WMH) prediction in unseen medical imaging datasets.

Main Methods:

• Focus on Domain Generalization (DG) where models are trained on source distributions without knowledge of the target distribution.
• Investigated the synergy between domain adversarial learning and mix-up techniques.
• Applied and evaluated these methods on the multi-site WMH Segmentation Challenge dataset and an in-house dataset for WMH prediction.

Main Results:

• Identified theoretical synergy between domain adversarial learning and mix-up for DG.
• Demonstrated significant improvements in white matter hyperintensity prediction on an unseen target domain.
• Showcased the effectiveness of the combined DG approaches in handling distribution shifts in medical imaging.

Conclusions:

• The proposed DG approach significantly enhances the generalizability of machine learning models in medical imaging.
• Combining domain adversarial learning and mix-up offers a promising strategy for robust WMH prediction across different data distributions.
• This work contributes to developing more reliable AI tools for diverse clinical settings.