No abstract

BackgroundPlain pelvic radiographs are the most common imaging modality used in the first line for diagnosis of axial Spondyloarthritis (axSpA). However, evaluation of the sacroiliac joint on two-dimensional plain radiographs may lead to misinterpretation among the evaluators.ObjectivesTo investigate the diagnostic power of deep learning models in conventional radiographs of the patients with axSpA.MethodsThe study included 320 axSpA patients and 348 healthy controls (age; 34.5±15.3/38.2±10.6, p=0.072; gender (male) 53.8%/46.3%, p=0.064). Sacroiliitis was confirmed on sacroiliac MRI according to the Assessment of Spondyloarthritis International Society (ASAS) definition. The contrast equalization was preprocessed with the Clahel (Contrast-Limited Adaptive Histogram Equalization) filter. Then, classification was performed with Alexnet, VGG16, resnet101 and resnet50 models. As a result of the trials, the best result was achieved with resnet50. Four different filtering scenarios were applied (Clahel filter cliplimit 0.25, Clahel filter cliplimit 0.50, clahel filter 1.00 and no filter).Two different cropping processes were performed on the direct radiographs, and uncropped, cropped at pelvic borders, cropped images close to the sacroiliac joint were applied to the deep learning model. Meanwhile, all images were also evaluated by 3 rheumatologists for the presence of sacroiliitis.ResultsAccording to the results of 4 different scenarios studied with the Resnet50 model, the best result was obtained with the RESNET50 Model + Clahel filter clipLimit 0.50. With this model, after applying the clahel filter with a coefficient of 0.5 to the full resolution data, we achieved 0.8135 success in the separation of AS and normal. A kappa error of 0.0561, Cohen’s Kappa Error = 0.6267, Fscore 0.8022 (AS), 0.8253 (normal) were obtained. After applying the clahel filter with a coefficient of 0.5 with Resnet50 to the pelvic data, we achieved 0.625 success in separation of AS and normal. Kappa error 0.0694, Cohen’s Kappa Error=0.2400, Fscore 0.5399 (AS), 0.6838 (normal) values ​​were obtained. After applying the Clahel filter with a coefficient of 0.5 with Resnet50 to the sacroiliac data, we achieved 0.61 success in the separation of AS and normal. Kappa error 0.0696, Cohen’s Kappa Error=0.2131, Fscore 0.5517 (AS), 0.6549 (normal) values ​​were obtained. As a result of the evaluation of the radiographs by the clinician, Cohen’s kappa was 0.452 and accuracy was 0.73 for the first rheumatologist; Cohen’s kappa 0.132 and accuracy 0.56 for the second rheumatologist and Cohen’s kappa 0.362 and accuracy 0.68 for the third rheumatologist were found.ConclusionApplication of RESNET50 Model + Clahel filter (‘clipLimit’, 0.5, ‘Distribution’, ‘rayleigh’) on uncropped images showed higher precision in diagnosing sacroiliitis from conventional radiographs compared to other filtering scenarios. Our results were found to have higher accuracy than the evaluation of three rheumatologists.Table 1.Precision and kappa values of Resnet50 model, 4 different filtering scenarios and 2 different clipping operationsAccuracyCohen’s kappa (CI)*RESNET50+no filter0.73060.4509 (0.3234-0.5785)*RESNET50 Model+ Clahel filter (0,25)0.69430.3889 (0.2590-0.5189)*RESNET50 Model+ Clahel filter (0,50)0.81350.6267 (0.4186-0.7555)*RESNET50 Model+ Clahel filter (1,00)0.68390.3552 (0.2213-0.4890)**RESNET50 Model+ Clahel filter (0,50)0.6250.2400 (0.1523-0.3245)***RESNET50 Model+ Clahel filter (0,50)0.6100.2131 (0.1121-0.3430)* Uncropped sacroiliac X-Rays; ** Cropped X-Rays at the pelvic margins; *** Cropped X-Rays close to the sacroiliac jointsFigure 1.Steps of deep learning model in X-Ray imagesDisclosure of InterestsNone declared