SAT0564 BONE TEXTURE ANALYSIS WITH DEEP LEARNING IN HAND RADIOGRAPHS FOR ASSESSING THE RISK OF RHEUMATOID ARTHRITIS. (2nd June 2020)

• Record Type: Journal Article
• Authors: Kuo, C. F.; Miao, S.; Zheng, K.; Lu, L.; Hsieh, C. I.; Lin, C.
• Is Part Of: Annals of the rheumatic diseases. Volume 79(2020)Supplement 1
• Page Start: 1240
• Page End: 1240
• Publication Date: 2020-06-02
• Subjects: Rheumatism -- Periodicals; 616.723005
• DOI: 10.1136/annrheumdis-2020-eular.837
• Languages: English
• ISSNs: 0003-4967

Abstract

Abstract : Background: Conventional x-rays are essential to identify radiographic changes of rheumatoid arthritis (RA) in structure and bone texture. Limited evidence suggests that the bone texture analysis may quantify the radiographic changes in RA; 1 however, current techniques such as the fractal dimension characterize fixed texture features. Deep learning offers novel methods to 'learn' radiographic texture features relevant to RA. Objectives: To develop a deep learning model to assess the radiographic bone texture in the distal metacarpal bone relevant to RA. Methods: We collected 3, 738 conventional hand radiographs from 2, 128 individuals (RA, n = 908; non-RA, n = 1220). The second, third, and fourth metacarpal bone images were segmented using a curve Graph Convolutional Network (GCN), and the distal third was used as the input to train a texture model to classify RA. The texture model was based on the Deep Texture Encoding Network (Deep-TEN) architecture (figure 1 ), 2 which put an encoding layer on top of a pre-trained 18-layered residual network (ResNet18). The vectors produced by the model represent the orderless texture features that were used to generate a texture score for RA. Five texture models are trained using 5-fold cross-validation and are ensembled during inference by averaging the model outputs to produce the final score. We then validate the model using hand radiographs of 166 RA patients and 166 non-RA patients. Overall model performance was measured by area under the curve of the receiver operator curve (AUROC). Multivariate logistic regression was used to estimate the odds ratio (OR) and 95% confidence interval (CI) of RA. Results: We included 140 women and 26 men with RA (mean age, 55.9±1.8 years) and 166 non-RA individuals (F: M, 140:26; mean age, 55.5 ± 1.8 years). The mean texture score was 0.49 (95% CI, 0.48–0.50) in RA patients, which is significantly higher than non-RA patients (0.42, 95% CI, 0.40–0.43; p<0.01). The AUROC of the model was 0.68. In the multivariate logistic regression model, a high texture score (>0.43) is associated with an OR (95% CI) of 3.42 (2.48–4.72) for RA, adjusted by age and sex. Conclusion: This study indicates that the texture model can delineate radiographic changes in texture relevant to RA and, coupled with automatic joint detection and segmentation, it has the potential to aid early RA diagnosis and monitor radiographic progression. References: [1]Zandieh S, Haller J, Bernt R, et al. Fractal analysis of subchondral bone changes of the hand in rheumatoid arthritis. Medicine (Baltimore) 2017;96(11):e6344. [2]Zhang H, Xue J, Dana K. Deep TEN: Texture Encoding Network. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017:708-17. Disclosure of Interests: None declared