Email Alert | RSS

Chinese Journal of Antituberculosis ›› 2026, Vol. 48 ›› Issue (8): 1184-1191.doi: 10.19982/j.issn.1000-6621.20250524

• Original Articles • Previous Articles     Next Articles

Multi-center study on CE-T1WI radiomics-based machine learning model in differentiating tuberculous spondylitis from brucella spondylitis

Guo Xiaowen1, Jiao Tianyu1, Sun Li1, Ni Conghui1, Wang Zilin2, Wang Hao1()   

  1. 1 Department of Radiology, Public Health Clinical Center Affiliated to Shandong University, Ji’nan 250100, China
    2 Department of Medical College Affiliated to Nanchang University, Nanchang 330006, China.
  • Received:2025-12-24 Online:2026-08-01 Published:2026-07-30
  • Contact: Wang Hao,Email:wh630405@163.com

Abstract:

Objective: To evaluate the value of a contrast-enhanced T1-weighted imaging (CE-T1WI) based radiomics model in differentiating tuberculous spondylitis (TS) from brucella spondylitis (BS). Methods: Clinical data and CE-T1WI images from 274 confirmed patients (131 TS, 143 BS) in 4 medical centers were collected. Radiomics features were extracted from manually delineated regions of interest (ROIs) using 3D-Slicer. Feature selection was performed using the Mann-Whitney U test (P<0.05), Spearman correlation analysis (coefficient>0.9), and the LASSO algorithm. The dataset was randomly split into a training cohort (n=219) and a validation cohort (n=55) in an 8∶2 ratio. Four machine learning models (Random Forest, Extra Trees, XGBoost, LightGBM) were established, and their diagnostic performances were evaluated using ROC curves. Results: A total of 107 radiomics features were initially extracted from CE-T1WI images. After feature selection, 11 optimal features were retained for model construction, including 7 texture features, 3 shape features, and 1 first-order feature. In the validation cohort, the AUC values of Random Forest, Extra Trees, XGBoost, and LightGBM models established based on these features were 0.773, 0.822, 0.838, and 0.702, respectively. Among them, the XGBoost model demonstrated the best overall diagnostic performance with relatively high accurate rate, sensitivity and specificity. Conclusion: The CE-T1WI-based radiomics machine learning models are effective for differentiating TS from BS. The XGBoost model showed best diagnostic performance, making it a promising tool for clinical decision-making.

Key words: Magnetic resonance imaging, Artificial intelligence, Tuberculosis, spinal, Brucellosis

CLC Number: