Diagnosing Ankylosing Spondylitis via Architecture-Modified ResNet and Combined Conventional Magnetic Resonance Imagery.

Abstract

Ankylosing spondylitis (AS), a lifelong inflammatory disease, leads to fusion of vertebrae and sacroiliac joints (SIJs) if undiagnosed. Conventional magnetic resonance imaging (MRI), e.g., T1w/T2w, is the diagnostic modality of choice for AS. However, computed tomography (CT)-a second-line modality-offers higher specificity because CT differentiates AS-relevant bony erosions/lesions better than MRI. We wished to ascertain whether MRI could be used to train/optimize convolutional neural networks (CNNs) for AS classification and which type of conventional MRI may dominate. We extracted 534 AS and 606 control SIJs from 56 patients with three simultaneously captured conventional MRI sequences. For classification, we compared modified/optimized variants of ResNet50, InceptionV3, and VGG16. CNNs were fine-tuned using 6-fold cross-validation and optimized architecturally and by learning rate. To automate SIJ extraction, we also developed a YOLOv5-based SIJ detector. Models trained on images that were the RGB combination of the MRI sequences significantly outperformed models trained on any one sequence ( $p<0.05$ ). The best architecture, located via architectural decomposition, was the first 9 blocks of ResNet50. The reduced-parameters model, which met or exceeded the full architecture's performance in 83% less parameters, achieved a cross-validation test set accuracy, sensitivity, specificity, and ROC AUC of 95.26%, 96.25%, 94.39%, and 99.1%. Our SIJ detector achieved 96.88-99.88% mAP@0.5. Deep learning models successfully diagnose AS from control SIJs. Models trained on combined conventional MRI achieve high sensitivity and specificity, mitigating the need for radioactive CT.