Deep learning algorithm for semiquantification of spinal inflammation in axial spondyloarthritis

Objective To develop a deep learning algorithm for semiquantification of spinal inflammation in patients with axial spondyloarthritis (SpA). Methods The study included 330 participants with axial SpA. All patients underwent whole spine MRI with short τ inversion recovery (STIR) sequence by 3T MR unit. Three independent readers identified regions of interest to locate bone marrow oedema (BMO) and performed Spondyloarthritis Research Consortium of Canada (SPARCC) scoring. Two deep learning models based on attention Unet were developed. The BMO model differentiated image with or without spinal inflammation. The vertebral body (VB)-intervertebral disc (IVD) model identified discovertebral units for localisation. The intraclass correlation coefficient (ICC) and Pearson coefficient were used to evaluate agreement and correlation between scorings by human readers and deep learning-based pipeline. Performance of the models was evaluated using sensitivity, specificity, accuracy and Dice coefficient. Results The ICC and the Pearson coefficient of SPARCC scores between human readers and the deep learning-based scoring pipeline were 0.80 and 0.82, respectively. The sensitivity and specificity of spinal inflammation identification were 0.90 and 0.84, respectively. The Dice coefficients were 0.81 (VB) and 0.80 (IVD) in images with spinal inflammation. Conclusion The high consistency of the scoring pipeline with human readers suggested that the deep learning-based algorithm has the potential to provide semiquantitative assessment of spinal inflammation based on SPARCC in axial SpA.