BiNETR:MRI skull segmentation method based on bi-stream pyramid decoder and deep supervision
10.3969/j.issn.1005-202X.2025.08.006
- VernacularTitle:BiNETR:基于双流金字塔解码器和深监督的MRI颅骨分割方法
- Author:
Hongzhu WU
1
;
Xiaolin LI
;
Bo PENG
;
Zhiyong ZHOU
;
Yakang DAI
Author Information
1. 徐州医科大学医学影像学院,江苏 徐州 221004
- Publication Type:Journal Article
- Keywords:
skull segmentation;
deep learning;
bi-stream pyramid;
magnetic resonance imaging;
deep supervision
- From:
Chinese Journal of Medical Physics
2025;42(8):1018-1025
- CountryChina
- Language:Chinese
-
Abstract:
Skull segmentation in magnetic resonance image(MRI)provides realistic skull models for MEG and EEG positive problems.An MRI skull segmentation method based on bi-stream pyramid decoder and deep supervision(BiNETR)is proposed to solve the problem of difficult segmentation due to the blurred and complex structure of MRI skull imaging.The method uses a bi-stream pyramid decoder as the main decoder in the network structure of encoding-decoding,including serial dual decoders for edge information oriented and precise feature merging.Specifically,edge information oriented pyramid decoder effectively enhances the edge information based on feature sharpening to improve the edge segmentation accuracy,and the precise feature merging pyramid decoder further refines and reuses the edge-enhanced features to promote the fusion of deep and shallow features.In addition,deep supervised computation of intermediate feature loss is introduced to implant the gradient into the deep network for enhancing network training.The segmentation algorithm is validated on the skull dataset,achieving a Dice similarity coefficient of 0.880±0.039 and an average symmetric surface distance of(0.931±0.286)mm,outperforming other state-of-the-art methods.The experimental results demonstrate the effectiveness and accuracy of the proposed algorithm in MRI skull segmentation.
