Deep learning-based segmentation method of neck skeletal muscle in radiotherapy patients with head and neck tumors
10.19745/j.1003-8868.2025137
- VernacularTitle:基于深度学习的头颈部肿瘤放疗患者颈部骨骼肌分割方法研究
- Author:
Zhi MING
1
;
Ke LIU
;
Bin ZENG
;
Zhe WU
;
Mu-jun LIU
Author Information
1. 自贡市第一人民医院肿瘤科,四川 自贡 643000
- Publication Type:Journal Article
- Keywords:
deep learning;
head and neck tumor;
skeletal muscle;
intelligent segmentation;
UNet;
Mamba architecture;
attention mechanism
- From:
Chinese Medical Equipment Journal
2025;46(8):11-17
- CountryChina
- Language:Chinese
-
Abstract:
Objective To propose a lightweight deep learning network-based segmentation method for automatic segmenta-tion of the skeletal muscle at the third cervical spine(C3)level.Methods Firstly,121 patients with head and neck tumors admitted to the Department of Oncology of Zigong First People's Hospital from January 2019 to December 2022 were selected and randomly divided into a training set,a validation set and a test set in the ratio of 7∶1∶2.Secondly,a lightweight Mamba architecture was introduced into the UNet network and an attention gate(AG)mechanism was added to the skip connection path to construct a MB-UNet network model.Finally,the trained network models were evaluated for segmentation performance on the test set,the MB-UNet network model was compared with manual segmentation over the results of determination of skeletal muscle area(SMA),and with classical network models in terms of parameter scale and computation effort including UNet,Deeplab V3+,U2Net,VMUNet and UltraLight-VMUNet models.The time required by the MB-UNet network model for predicting SMA and that by the physician with the assistance of the model was summarized.Results When used for segmenting the skeletal muscle at C3 level the constructued MB-UNet network model gained advantages over the classical models,with a Dice similarity coefficient of 88.23%,an intersection over union(IoU)ratio of 78.94%,a sensitivity of 91.27%and a 95%Hausdorff distance of 7.13 mm;the SMA determined by manual segementation was basically close to that by the MB-UNet network model;the MB-UNet network model behaved generally better than the classical network models,with the computation effort being 1.88 GFLOPS and the parameter scale being 0.77M;it took the MB-UNet network model 1.93 s for the prediction on the test set,and only 2 min for the physician to obtain satisfactory results with the assistance of the MB-UNet network model,which was significantly shorter than that by munual segmentation(20 min).Conclusion The proposed method contributes to segmenting the skeletal muscle at C3 level precisely and rapidly and calculating SMA accurately,which helps clinicians to quickly diagnose sarcopenia in patients with head and neck tumors and improves the diagnostic efficiency.[Chinese Medical Equipment Journal,2025,46(8):11-17]
