Pan-Contrast Thalamic Nuclei Segmentation: Physics-Informed Image Synthesis for Performance Across All Contrast Variations

Motivation: Few tools exist for segmenting thalamic nuclei, and the ones available are limited to MR images captured with specific sequences. Goal(s): Develop a pan-contrast thalamic nuclei segmentation tool. Approach: Utilize a comprehensive physics-informed MR image synthesis platform to generate thalamus-enhanced MR images and train a U-Net for segmenting thalamic nuclei across various MR contrasts. Results: Thalamus-enhanced custom images boost inter-nuclei contrast by up to 58%. This enhancement improves state-of-the-art labeling accuracy by nearly 11% and facilitates the training of UTN, a thalamic nuclei segmentation network that achieves a pan-contrast median segmentation dice score of 0.816. Impact: UTN is the first thalamus segmentation tool that works across different contrast types, matching the performance of top tools, which are limited to white matter-nulled images. This allows for volumetric thalamic nuclei studies with any type of MR data.