Classification of radiographic and non-radiographic axial spondylarthritis in pelvic radiography using deep convolution neural network models

Discriminating radiographic axial spondyloarthritis (r-axSpA) from nonradiographic axial spondyloarthritis (nr-axSpA) using pelvic radiographs is challenging, especially for inexperienced clinicians. This study aims to perform deep convolution neuronal network (CNN) models to aid in this diagnostic challenge by using their radiographs. Six-hundred sacroiliac joint exams (300 pelvic radiographs) of patients from axSpA cohort were enrolled (screened between Jan 2010 and Jan 2020). All radiographs were examined and graded by a blinded expert rheumatologist. Four CNN models (VGG16, ResNet, DenseNet, and MobileNet) were proposed by combining them with the YOLOv7 object detection algorithm to mark the sacroiliac joints. The classification results of CNNs were evaluated by performance metrics [accuracy, AUROC (area under the receiver operating characteristic curve)]. The VGG16 model with the YOLOv7 algorithm yielded the best performance [accuracy of 83.8% (95% CI; 73.3-92.9%)]. The accuracy values of other models were 70.7% (58.3-82.7%), 77.1% (65.1-87.3%), and 71.8% (59.0-83.1%) for ResNet, DenseNet, and MobileNet, respectively. In the ROC analysis, the AUC value of the VGG16 model (AUC = 0.882) was higher than other CNNs (AUCs = 0.836, 0.808, and 0.787; DenseNet, ResNet, and MobileNet, respectively). This paper revealed deep learning architectures were able to differentiate r-axSpA from nr-axSpA on pelvic radiographs. Hereby, these models might be used as a clinical decision support system in clinical practice.