The Journal of Practical Medicine ›› 2026, Vol. 42 ›› Issue (15): 2786-2797.doi: 10.3969/j.issn.1006-5725.2026.15.015

• Treatise:Mechanism and Practice • Previous Articles    

Development of a disease activity assessment model for rheumatoid arthritis using ultrasound radiomics and deep learning

Xin ZHU1,2,Yangbiao LIANG2,Xiong XIAO1,2,Jialin XIE1,2,Junyan ZHU1,2,Yan CHEN2()   

  1. 1.The Second School of Clinical Medicine,Southern Medical University,Guangzhou 510515,Guangdong,China
    2.Department of Ultrasonography,Zhujiang Hospital,Southern Medical University,Guangzhou 510282,Guangdong,China
  • Received:2026-04-15 Revised:2026-06-04 Accepted:2026-06-08 Online:2026-08-10 Published:2026-08-13
  • Contact: Yan CHEN E-mail:smu_chen@163.com

Abstract:

Objective To develop and preliminarily evaluate the value of an integrated ultrasound model that combines radiomics and deep learning features for the objective assessment of disease activity in rheumatoid arthritis (RA). Methods A total of 184 affected joints from 92 RA patients were included and then divided into training and testing sets at an 8:2 ratio. After the delineation of the region-of-interest (ROI), the extraction and screening of features, and the cropping of joint cross-sectional images, three standalone models, namely ultrasound scoring, bioinformatics, and deep learning, were constructed. Additionally, two combined models, ultrasound-bioinformatics and ultrasound-bioinformatics-deep learning, were also built. The performance of the models was evaluated using receiver operating characteristic (ROC) curves, the area under the curve (AUC), and decision-making analysis (DCA) curves. Results In the test set, when compared with the ultrasound scoring model based on single features (AUC = 0.832, 95% CI: 0.709 - 0.954), the radiomics model (AUC = 0.824, 95% CI: 0.701 - 0.947), or the deep learning feature model (AUC = 0.788, 95% CI: 0.642 - 0.933), the combined ultrasound-radiomics-deep learning model (AUC = 0.901, 95% CI: 0.815 - 0.988) exhibited the best performance. The Hosmer-Lemeshow goodness-of-fit test for the combined ultrasound-radiomics-deep learning model yielded a p-value greater than 0.05, indicating adequate model calibration. Moreover, DCA showed that the combined model provides a higher net clinical benefit within a defined range of threshold probabilities. Conclusion This multimodal fusion quantification model exhibits excellent efficacy in evaluating the activity of RA disease, establishing a methodological basis for the formulation of future stratified treatment strategies and dynamic efficacy monitoring for RA.

Key words: rheumatoid arthritis, ultrasound, radiomics, deep learning

CLC Number: