Detection of sacroiliitis in spondyloarthritis patients via deep learning from computed tomography scans

Objective: Ankylosing spondylitis (AS) typically begins in young adulthood, and early diagnosis is associated with improved quality of life.Imaging-based assessment of sacroiliitis is essential for the diagnosis of AS.We developed and validated a deeplearning algorithm for grading sacroiliitis using computed tomography (CT).Methods: A deep learning model was developed using 1,674 CT scans from patients presenting with back pain at a tertiary hospital between January 2012 and June 2023.The sacrum and pelvic bones were segmented, and CT images were divided into left and right patches centered on the sacrum.Three-dimensional patches were converted into digitally reconstructed radiograph (DRR) images, with red, green, and blue channels assigned by layering depth information.Each patch was augmented into eight variants by adjusting a virtual cubic bounding box encompassing the sacroiliac joints.A ConvNeXt-Large model pretrained on ImageNet was used to grade sacroiliitis from each DRR view.Results: The 2.5D convolutional neural network demonstrated robust performance for sacroiliitis grading, with area under the receiver operating characteristic curve values for classes 0, 1, and 2 of 0.96 (95% confidence interval [CI], 0.95~0.98),0.84 (95% CI, 0.77~0.90),and 0.97 (95% CI, 0.95~0.98) in the internal test set and 0.96 (95% CI, 0.93~0.98),0.55 (95% CI, 0.46~0.64),and 0.92 (95% CI, 0.88~0.95) in the external test set, respectively.Conclusion: This approach demonstrated strong performance in grading sacroiliitis using CT images acquired with various imaging protocols.It may facilitate the early identification of sacroiliitis on routine CT examinations.