Development and validation of a deep learning-based low-dose cervical spine X-ray segmentation model
10.3969/j.issn.1002-1671.2025.07.032
- VernacularTitle:基于深度学习的低剂量颈椎X线分割模型构建与验证
- Author:
Zhenbo CHEN
1
;
Hongxia ZHANG
1
;
Weiyong YU
1
;
Xinying CONG
1
;
Tian ZHANG
1
;
Yang XIE
1
Author Information
1. 中国康复研究中心(北京博爱医院)影像科,北京 100068;首都医科大学康复医学院,北京 100068
- Publication Type:Journal Article
- Keywords:
cervical segmentation;
X-ray photography;
low-dose;
deep learning
- From:
Journal of Practical Radiology
2025;41(7):1225-1229
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop and validate a deep learning-based segmentation model for low-dose cervical spine X-ray,aiming to address the insufficient segmentation accuracy in low-dose protocols while balancing radiation protection and diagnostic accuracy.Methods A total of 1 363 patients cervical spine X-ray images data were collected.A dose-attenuation mathematical simulation sys-tem was constructed to generate 14 122 dynamic low-dose cervical spine images incorporating quantum noise,contrast degradation,and blur artifacts.A neural network model was developed for automated segmentation of low-dose cervical spine X-ray using this dataset.Results Within the dose range of 50%to 7.5%,the average reults of automatic segmentation by the neural network model and manual segmentation for each group were as follows:50%dose group,intersection over union(IoU)=0.98 vs 0.93(P=0.707)and Dice coefficient(Dice)=0.99 vs 0.96(P=0.749);10%dose group,IoU=0.97 vs 0.87(P=0.201)and Dice=0.99 vs 0.93(P=0.219);7.5%dose group,IoU=0.97 vs 0.67(P<0.01)and Dice=0.98 vs 0.80(P<0.01).Conclusion The developed deep learning model achieved robust cervical spine segmentation(IoU>0.96,Dice>0.98)below diagnostic dose thresholds[peak signal-to-noise ratio(PSNR)<38 dB,structural similarity index(SSIM)<0.65].Under ultra-low-dose conditions(PSNR=27.710 dB,SSIM=0.274),it demonstrated a 44.78%IoU improvement and 22.5%Dice improvement over manual segmentation.This model enables minimal radia-tion exposure while preserving diagnostic performance,confirming its theoretical feasibility for low-dose X-ray image analysis and clinical research potential.
