Automatic recognition and segmentation of brachial plexus in ultrasonic images based on deep learning
10.3760/cma.j.cn131148-20250216-00076
- VernacularTitle:基于深度学习的超声图像中臂丛神经的自动分割与识别
- Author:
Duo SHI
1
;
Han ZHANG
;
Peipei LIU
;
Ruichao ZHANG
;
Qingyu LIU
;
Hao SUN
;
Xiaofang FU
;
Mengjie DOU
;
Junpu HU
;
Changqin SUN
;
Keyan LI
;
Jianqiu HU
;
Guangquan ZHOU
;
Ligang CUI
;
Ping ZHOU
;
Faqin LYU
Author Information
1. 锦州医科大学解放军总医院第三医学中心研究生培养基地，北京 100039
- Publication Type:Journal Article
- Keywords:
Ultrasonography;
Deep learning;
Ultrasound imaging;
Brachial plexus
- From:
Chinese Journal of Ultrasonography
2025;34(9):737-744
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To propose a deep learning（DL）-based ultrasound imaging auxiliary tool for automatic segmentation and recognition of the brachial plexus（BP），and to enhance the accuracy and safety of clinical procedures.Methods:It was a multicenter study that collected 773 healthy subjects from Peking University Third Hospital and its branch campuses，the Third Medical Center of the Chinese PLA General Hospital，and Shanghai Eighth People's Hospital between August 2024 and February 2025. Brachial plexus（BP）images in the interscalene groove were captured used high-frequency ultrasound by senior sonographers，a dataset comprising 1 289 standardized images were constructed and the improved model（CHA-TransUNet）was trained. The test set was input into 6 different models（CHA-TransUNet，R50-Unet，TransUnet，SegFormer，SwinUnet，MISSFormer）for segmentation. Segmentation accuracy was evaluated using metrics including the Dice similarity coefficient（DSC），95% Hausdorff distance（HD95）and mean intersection over union（mIoU），and was compared with the segmentation results of 3 ultrasound physicians with varying experience levels（junior physicians and senior physicians）to validate the model's segmentation efficacy.Results:The CHA-TransUNet model established based on a dataset of 1 289 standardized images achieved segmentation results for the BP with a DSC of 90.15%，mIoU of 91.02%，and HD95 of 8.08. Its accuracy was higher than other mainstream models（DSC：90.15% vs. 87.60%，87.77%，81.35%，84.78%，84.55%），significantly better than junior physicians（DSC：90.15% vs. 68.73%， Z=-127.76， P<0.001），and approached the level of senior physician（DSC：90.15% vs. 86.15%， Z=-31.33， P=0.549）. The model demonstrated superior boundary recognition in complex anatomical structures（e.g.，C6/C7 nerve roots）compared to ultrasound physicians（junior and senior）（HD95：8.08 vs. 26.34，17.44，56.80）. Conclusions:This study proposes an analysis model for BP ultrasound images，CHA-TransUNet. This model achieves segmentation and recognition of the BP with relatively complex pathways and structures. The model exhibits high accuracy and stability，outperforming current mainstream network models and junior physicians while approaching the performance level of senior physicians. It assists junior physicians or trainees in more accurately identifying and localizing the BP.
