Improving Substantia Nigra Segmentation Across Different Neuromelanin-Sensitive MRI Sequences Using Domain Generalization Techniques

Motivation: Automatic segmentation of the substantia nigra (SN) in neuromelanin-sensitive MRI (NM-MRI) is challenging due to variations across NM-MRI sequences. Goal(s): To enhance SN segmentation using domain generalization techniques, creating a domain-agnostic segmentation model robust to variation in sequence parameters. Approach: A U-Net model was trained on an individual sequence and tested on unseen ones. Data augmentation techniques were applied to address structural and intensity variations across fundamental sequence differences. Results: Preliminary results suggest that while data augmentation can improve SN segmentation across different sequences, robust deep-learning-based segmentation remains challenging with sequence effects hindering the model's ability to automatically delineate SN across unseen parameters. Impact: Applying data augmentation techniques significantly enhances automated substantia nigra segmentation in neuromelanin-sensitive MRI, advancing the development of robust, clinically reliable models adaptable to various imaging methods and neurodegenerative conditions.