Related Experiment Video
Updated: Mar 13, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Automated segmentation and quantitative measurement of cervical nerves in ultrasound images using an SZJ-SEG-based
Zheyuan Zhang1, Wen Cao2, Haomei Luan1
1Department of Ultrasound, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua Medicine, Tsinghua University, Beijing, China.
Background:
Accurate recognition and quantitative assessment of cervical nerves are essential for ultrasound-guided nerve block, preoperative evaluation, and diagnosis of peripheral neuropathies. However, manual interpretation and measurement of nerve structures remain time-consuming and operator dependent. This study aimed to develop an intelligent, fully automated system for precise segmentation and quantitative measurement of cervical nerves in ultrasound images using a deep learning-based approach.
Methods:
Ultrasound images from 200 healthy volunteers were collected and meticulously annotated to construct a large-scale standardized dataset comprising 117,729 images. A fully automated analysis framework was designed, incorporating a YOLOv11 network for precise localization of the effective imaging region [mean intersection over union (mIoU) =0.99] and an optical character recognition (OCR) module for automatic extraction of depth scale information. The proposed segmentation network, termed SZJ-SEG, combines a Deconv Block and an efficient upsampling convolution block (EUCB) based on a ResNet50 backbone to achieve high segmentation accuracy with low computational cost. Quantitative analysis of the C5-C7 nerve roots was performed through pixel-to-physical scale calibration for cross-sectional area (CSA) and perimeter calculation.
Results:
On a clinical dataset comprising 117,729 images (training/validation/test =94,183/11,773/11,773), SZJ-SEG achieved mIoU values of 0.9124, 0.9109, and 0.9041 for the C5, C6, and C7 nerves, respectively, and 0.9227 for the continuous brachial plexus section. The mean absolute error (MAE) of CSA measurement ranged from 0.278 to 0.442 mm2, with mean absolute percentage error (MAPE) between 2.43% and 6.16%, and Pearson correlation coefficients exceeding 0.96. For perimeter measurements, MAE ranged from 0.374 to 0.471 mm, MAPE from 3.05% to 4.49%, and Pearson R from 0.84 to 0.91, indicating excellent consistency with manual annotations.
Conclusions:
The proposed SZJ-SEG-based system enables accurate, efficient, and reproducible segmentation and quantitative analysis of cervical ultrasound images. With strong performance and high clinical relevance, it provides a reliable tool for ultrasound-guided nerve block localization and preoperative assessment. Its modular architecture offers scalability for extension to other anatomical regions, highlighting broad potential in intelligent ultrasound diagnosis.
