Radiographic bone texture analysis using deep learning models for early rheumatoid arthritis diagnosis (Preprint)

• DOI: 10.2196/preprints.48350
• Publication Date: 2023-04-26T16:28:18Z
• Authors: Yun-Ju Huang; Chiung-Hung Lin; Shun Miao; Kang Zheng; Le Lu; Yuhang Lu; Chihung Lin; Chang-Fu Kuo
• Keywords: Texture (cosmology); Area under curve

<sec> <title>BACKGROUND</title> Rheumatoid arthritis (RA) is distinguished by the presence of modified bone microarchitecture, also known as 'texture,' in periarticular regions. The radiographic detection such alterations RA can be challenging. </sec> <title>OBJECTIVE</title> To train and to validate a deep learning model quantitatively produce texture features di-rectly from radiography predict diagnosis early without human reading. Two kinds models were compared for diagnostic performance. <title>METHODS</title> Anterior-posterior bilateral hands radiographs 891 (within one year initial diagno-sis) 1237 non-RA patients split into training set (64%), validation (16%), test (20%). second, third, fourth distal metacarpal areas segmented Deep Texture Encod-ing Network (Deep-TEN; texture-based) residual network-50 (ResNet-50; structure-based) probability RA. <title>RESULTS</title> area under curve receiver operating characteristics was 0.69 Deep-TEN 0.73 ResNet-50 model. positive predictive values high score classify using 0.64 0.67, respectively. High mean tex-ture scores associated with age- sex-adjusted odds ratios (ORs) 95% confidence interval (CI) 3.42 (2.59–4.50) 4.30 (3.26–5.69) models, moderate risk groups determined adjusted ORs (95% CIs) 2.48 (1.78–3.47) 4.39 (3.11–6.20) RA, respectively, those 2.17 (1.55–3.04) 6.91 (4.83–9.90), <title>CONCLUSIONS</title> Fully automated quantitative assessment help classification