Post-MI: unsupervised brain tissue segmentation via post-maximized mutual information.
Authors
Affiliations (8)
Affiliations (8)
- Computer School, Beijing Information Science and Technology University, Beijing, China.
- Department of Pediatrics, Zhongshan Hospital Xiamen University, Xiamen, Fujian, China.
- Department of Neurology, The Third Hospital of Xiamen, Xiamen, Fujian, China.
- Department of Neurology and Geriatrics, Fujian Institute of Geriatrics, Fujian Medical University Union Hospital, Fuzhou, Fujian, China.
- Department of Neurology and Department of Neuroscience, The First Affiliated Hospital of Xiamen University, School of Medicine, Xiamen University, Xiamen, Fujian, China.
- Xiamen Medical Quality Control Center for Neurology, Xiamen, Fujian, China.
- Department of Ophthalmology, Peking University People's Hospital, Beijing, China.
- Beijing Key Laboratory of Ocular Disease and Optometry Science, Peking University People's Hospital, Beijing, China.
Abstract
Brain tissue segmentation in magnetic resonance imaging (MRI) is a fundamental step in quantitative neuroimaging and supports the analysis of structural brain alterations associated with aging. However, voxel-level annotation is expensive and labor-intensive, which limits the development of artificial intelligence methods in annotation-scarce neuroimaging settings. Existing mutual-information-based unsupervised segmentation methods mainly model consistency across different views of the same sample, while the representation derived from draft segmentation remains underexplored for mutual information modeling. In this paper, we propose Post-MI, an unsupervised brain tissue segmentation method that estimates mutual information between deep features extracted from the raw image and deep features derived from the probability map of the draft segmentation. A post-training strategy is introduced to enable this objective to refine the segmentation network after draft segmentation generation. In addition, edge information and a discriminative loss are incorporated to enhance boundary preservation and inter-class separability. We evaluate the proposed method primarily on IBSR-18 and further conduct exploratory experiments on BraTS2018. On IBSR-18, Post-MI achieves a mean intersection over union of 0.3635 and a mean Dice score of 0.4661, outperforming the compared unsupervised baselines. Ablation experiments further verify the positive contributions of edge information and discriminative loss. On BraTS2018, exploratory results indicate that more complex tumor structures and weaker boundaries remain challenging for the current method. Post-MI provides a feasible solution for brain MRI segmentation under annotation-scarce conditions and may facilitate quantitative structural brain analysis in aging-related neuroimaging research and other studies of age-related neurological conditions where large-scale manual delineation is difficult to obtain.