Pan-Contrast Learning of MRI Segmentation for Healthy and Anomaly Cases: Faithful to Tissue Properties and MR Physics

Motivation: Current MRI segmentation methods are limited to certain scan types or do not consider MR physics, which negatively affects segmentation quality in some cases. Goal(s): Learn and validate pan-contrast MRI segmentation. Approach: We introduce a framework that accurately reflects tissue properties and MR physics in generating images for all MR sequences using randomized scanning and noise parameters, to aid learning and validation of pan-contrast segmentation. Results: UltBrainNet outperforms the state-of-the-art in 87% of labels for conventional MR images. UBN generalizes across the human lifespan, including challenging neonate cases, and outperforms the SOTA in 100% of labels for low-resolution, variable-orientation, and pathology cases. Impact: UBN offers a comprehensive solution for consistent segmentation across all MR image contrasts, vendors, resolutions, sites, preprocessing methods, and age groups.