Brain glioma MRI segmentation method integrating visual state space and depthwise separable convolution(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2026年第6期
- Page:
- 756-765
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Brain glioma MRI segmentation method integrating visual state space and depthwise separable convolution
- Author(s):
- MA Fangfang1; ZHAO Mingyue1; SHI Yaning2; WANG Ping1
- 1. Air Force Medical Center, Beijing 100142, China 2. Department of Neurosurgery, the First Hospital Affiliated to Army Medical University, Chongqing 400038, China
- Keywords:
- Keywords: brain glioma MRI image segmentation VMamba visual state space depthwise separable convolution
- PACS:
- R318；R739.4
- DOI:
- DOI:10.3969/j.issn.1005-202X.2026.06.008
- Abstract:
- Abstract: Objective To propose a state space-enhanced context and boundary-aware segmentation network (SCB-UNet) for enhancing the segmentation accuracy of brain glioma MRI images and improving the models ability to cope with the global infiltrative nature and complex boundaries of tumors. Methods A VMamba module was introduced into the encoder to model global image dependencies via a visual state space mechanism, thereby strengthening the models perception of the overall tumor structure. Furthermore, a multi-scale visual state space feedforward network was established in the decoder to integrate global context with local multi-scale features, thus boosting the models ability to distinguish heterogeneous components inside the tumor. Finally, a depthwise separable convolution expansion module is constructed to aggregate local contextual information before feature upsampling, which improves the accuracy of detail recovery for complex boundaries and small lesions. Results Experiments on the public BraTS2019 and BraTS2023 datasets show that the average Dice similarity coefficients of the SCB-UNet across different tumor regions reached 85.94% and 83.33%, respectively, and the average 95% Hausdorff distances were 4.18 and 5.13 mm, respectively. In the ablation experiment on the BraTS2023 dataset, compared with U-Net baseline model, the SCB-UNet reduced the average 95% Hausdorff distance by 1.46 mm, further verifying the effectiveness of each module. Conclusion The SCB-UNet provides a feasible technical solution for the accurate segmentation of brain glioma MRI images and demonstrates promising potential for clinical application.
Last Update: 2026-06-26