Extending an Anatomy-Aware Framework for Unsupervised Segmentation of Multiple Sclerosis Lesions

Document Type

Article

Publication Date

2-6-2026

Publication Title

IEEE Access

Keywords

Lesions, Bayes methods, Brain modeling, Magnetic resonance imaging, Adaptation models, Image segmentation, Probabilistic logic, Computational modeling, Anatomy, White matter

Abstract

Traditional unsupervised lesion segmentation methods focus on modeling the healthy brain distribution, detecting lesions as deviations from it. However, relying exclusively on healthy feature learning often results in high false positive rates, particularly under domain shift. Although advanced foundation models and probabilistic approaches-incorporating semantic information or intensity priors-have improved performance on large, well-contrasted lesions such as tumors, their generalizability to more subtle and complex pathologies such as multiple sclerosis (MS) remains challenging. We propose a new framework that combines Bayesian modeling with deep learning to segment MS lesions without using manual annotation. The aim is to leverage the unsupervised modeling capability of the Bayesian framework, which adaptively refines a deep learning-generated lesion map using anatomical priors and MRI observations. The method explicitly models healthy brain anatomy using tissue probabilities and detects lesions as hyperintense outliers within white matter regions, which is consistent with clinical reasoning. Evaluated on the ISBI dataset, our approach outperforms baseline unsupervised methods, with a Dice score of 0.52, an area under the precision-recall curve (AUPRC) of 0.51, an intersection-over-union (IoU) of 0.36, a sensitivity of 0.52, and a specificity of 0.99. By integrating anatomical priors with statistical intensity modeling within a Bayesian framework, our approach reduces false positives, which are common in previous methods, mainly due to boundary ambiguity and intensity overlap between lesions and gray matter.

Volume

14

First Page

24942

Last Page

24953

Find It @ Sladen

Share

COinS