Background:Osteoarthritis is a degenerative disorder characterized by radiographic features of asymmetric loss of joint space, subchondral sclerosis, and osteophyte formation. Conventional plain films are essential to detect structural changes in osteoarthritis. Recent evidence suggests that fractal- and entropy-based bone texture parameters may improve the prediction of radiographic osteoarthritis.1In contrast to the fixed texture features, deep learning models allow the comprehensive texture feature extraction and recognition relevant to osteoarthritis.Objectives:To assess the predictive value of deep learning-extracted bone texture features in the detection of radiographic osteoarthritis.Methods:We used data from the Osteoarthritis Initiative, which is a longitudinal study with 4,796 patients followed up and assessed for osteoarthritis. We used a training set of 25,978 images from 3,086 patients to develop the textual model. We use the BoneFinder software2to do the segmentation of distal femur and proximal tibia. We used the Deep Texture Encoding Network (Deep-TEN)3to encode the bone texture features into a vector, which is fed to a 5-way linear classifier for Kellgren and Lawrence grading for osteoarthritis classification. We also developed a Residual Network with 18 layers (ResNet18) for comparison since it deals with contours as well. Spearman’s correlation coefficient was used to assess the correlation between predicted and reference KL grades. We also test the performance of the model to identify osteoarthritis (KL grade≥2).Results:We obtained 6,490 knee radiographs from 446 female and 326 male patients who were not in the training sets to validate the performance of the models. The distribution of the KL grades in the training and testing sets were shown in Table 1. The Spearman’s correlation coefficient was 0.60 for the Deep-TEN and 0.67 for the ResNet18 model. Table 2 shows the performance of the models to detect osteoarthritis. The positive predictive value for Deep-TEN and ResNet18 model classification for OA was 81.37% and 87.46%, respectively.Table 1Distribution of KL grades in the training and testing sets.KL grades01234TotalTraining set1089341.9%458218.7%611423.5%332012.8%7993.1%25,978Testing set247238.1%135320.8%169626.1%77511.9%1943.0%6,490Table 2Performance matrices of the Deep-Ten and ResNet18 models to detect osteoarthritisDeep-TENResNet18Sensitivity62.29%(95% CI, 60.42%–64.13%)59.14%(95% CI, 57.24%–61.01%)Specificity90.07%(95% CI, 89.07%–91.00%)94.09%(95% CI, 93.30%–94.82%)Positive predictive value81.37%(95% CI, 79.81%–82.84%)87.46%(95% CI, 85.96%–88.82%)Negative predictive value77.42%(95% CI, 77.64%–79.65%)76.77%(95% CI, 75.93%–77.59%)Conclusion:This study demonstrates that the bone texture model performs reasonably well to detect radiographic osteoarthritis with a similar performance to the bone contour model.References:[1]Bertalan Z, Ljuhar R, Norman B, et al. Combining fractal- and entropy-based bone texture analysis for the prediction of osteoarthritis: data from the multicenter osteoarthritis study (MOST). Osteoarthritis Cartilage 2018;26:S49.[2]Lindner C, Wang CW, Huang CT, et al. Fully Automatic System for Accurate Localisation and Analysis of Cephalometric Landmarks in Lateral Cephalograms. Sci Rep 2016;6:33581.[3]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