MRI-Based Framework for Ankylosing Spondylitis Diagnosis

Ankylosing spondylitis (AS) is a progressive inflammatory disease of the sacroiliac joints and spine, and its early inflammation can be detected only with magnetic resonance imaging (MRI). Manual detection of early inflammation on MRI is time-consuming, inconsistent, and subjective, which delays diagnosis and treatment. This paper presents an automated approach to detecting AS from MRI scans. The approach segments and classifies MRI images automatically and makes the classification more interpretable. The framework combines several deep learning models. An Attention U-Net segments the sacroiliac joints in the MRI scans. A hybrid convolutional neural network (CNN) and Transformer architecture then classifies each image as AS-positive or AS-negative and attempts to assign the disease stage (normal, early, moderate, or advanced). This architecture uses EfficientNet to extract detailed local features and Vision Transformer blocks to capture global context. Finally, gradient-weighted class activation mapping (Grad-CAM) gives a visual explanation of each decision. On 962 test images, the framework detected AS with 88.67% accuracy; stage grading remained limited, with 60.40% stage accuracy, and most early and moderate cases were predicted as advanced.