Boundary-aware and uncertainty-guided deep learning for multimodal magnetic resonance imaging segmentation of stroke lesions.
Authors
Affiliations (6)
Affiliations (6)
- College of Information Engineering, Sichuan Agricultural University, Ya'an, China.
- Department of Neurology, Ya'an People's Hospital, Sichuan Province, Ya'an, China.
- Tianjin Key Laboratory of Radiation Medicine and Molecular Nuclear Medicine, Institute of Radiation Medicine, Chinese Academy of Medical Sciences and Peking Union Medical College, Tianjin, China.
- NewIdea Yunshu Technology Co., Ltd., Chengdu, China.
- Tsinghua University, Beijing, China.
- Department of Neurology, The Affiliated Hospital, Southwest Medical University, Luzhou, China.
Abstract
Acute ischemic stroke is a major cerebrovascular disease, and accurate lesion segmentation is important for quantitative image analysis and subsequent clinical assessment. Multimodal magnetic resonance imaging, including diffusion-weighted imaging, apparent diffusion coefficient maps, and fluid-attenuated inversion recovery, provides complementary information for lesion delineation. However, lesion appearance may vary across sequences, and blurred boundaries or small scattered lesions can make multimodal feature fusion unstable. We developed MBGS-UNet, a multi-branch deep learning model for acute ischemic stroke lesion segmentation. Separate encoders were used to preserve sequence-specific information. A boundary-aware graph attention module was introduced at deeper feature levels to regulate cross-modal interaction in ambiguous regions, and an uncertainty-gated fusion module was used to balance graph-enhanced features with encoder features at the voxel level. Uncertainty-weighted boundary supervision and consistency learning were further incorporated to emphasize difficult regions during training. The model was evaluated on the ISLES-2022 dataset using an independent test set and label-free sliding-window inference. MBGS-UNet achieved a Dice score of 0.7208 on the independent test set, numerically higher than M2FNet (0.7068) and UMamba (0.7016) under the same evaluation protocol. However, case-wise paired statistical analysis did not provide sufficient evidence to support clear superiority over these strong baselines. Visual analysis showed that high-uncertainty responses were mainly concentrated around lesion boundaries and other error-prone regions, indicating a spatial relationship between predictive uncertainty and local segmentation difficulty. MBGS-UNet achieved competitive segmentation performance while providing interpretable uncertainty visualization. The uncertainty maps may offer useful cues for identifying ambiguous lesion boundaries and other regions that require closer review. Nevertheless, the observed numerical improvement should not be interpreted as conclusive evidence of superiority or clinical validation. Further evaluation on larger, more diverse, and external cohorts is required to establish broader robustness and clinical applicability.