Post-MI: unsupervised brain tissue segmentation via post-maximized mutual information

Post-MI is a new method for brain tissue segmentation in magnetic resonance imaging (MRI) without requiring manual data annotation. The method estimates mutual information between deep features extracted from the original image and features derived from the probability map of draft segmentation. It uses a post-training strategy to refine the segmentation network and incorporates edge information and discriminative loss to improve boundary preservation. On IBSR-18 test data, Post-MI achieved a mean Dice score of 0.4661 and mean Intersection over Union of 0.3635, outperforming compared unsupervised methods. The method provides a solution for brain MRI segmentation under annotation-scarce conditions and may support the analysis of structural brain changes associated with aging.