Improving Subcortical Segmentation in Brain MRI Using Knowledge Distillation to Enhance Robustness Against Motion Artifacts

Motivation: Accurate segmentation of subcortical brain regions in MRI is challenging due to motion artifacts that distort structural details and affect subsequent analyses, necessitating improved approaches for segmentation reliability. Goal(s): To develop a knowledge distillation framework that reduces the influence of motion artifacts on MRI segmentation, thus enhancing subcortical accuracy without the need for motion correction preprocessing. Approach: A teacher model trained on motion-free data guides a student model trained on motion-corrupted data, improving segmentation accuracy without complex motion correction. Results: The framework improved Dice Similarity Coefficients in subcortical regions, demonstrating enhanced segmentation performance and robustness on motion-corrupted data. Impact: This approach improves MRI segmentation of motion-corrupted data, supporting reliable subcortical analysis without complex preprocessing. It provides a method for cleaner, artifact-resistant segmentation, presenting significant applications both in neurodevelopmental and neurodegenerative disease research.