Segmentation of Brain Lesions Using Posterior Distributions Learned by Subspace-assisted Generative Model

Motivation: Deep learning shows great potential for brain lesion segmentation but poor generalization (due to limited training data) could lead to false positives. Goal(s): Our goal was to improve the segmentation accuracy by learning target-specific posterior distributions. Approach: We proposed a new Bayesian brain lesion segmentation method, leveraging posterior distributions learning, including both posterior normal and lesion distributions, through a subspace-assisted deep generative model. Results: The proposed method achieved significantly improved segmentation performance across multiple public datasets with stroke, tumor, and multiple sclerosis lesions, in comparison with the state-of-the-art methods. Impact: The proposed method significantly improved accuracy and robustness of lesions segmentation in brain MR images, which may provide a useful tool for brain lesion delineation in image processing and clinical applications.