High-resolution deep learning DIXON MRI of the sacroiliac joints is non-inferior to standard MRI in patients with suspected axial Spondyloarthritis.

Abstract

To compare the standard multi-sequence MRI protocol (sMRI) of the sacroiliac joints with a single high-resolution deep learning-reconstructed DIXON sequence (DL-DIXON) in patients with suspected axial spondyloarthritis (axSpA). Seventy-six patients with chronic low back pain and suspected axSpA underwent clinical, laboratory, and genetic assessment followed by 3T sMRI (T1, T2-FS, VIBE, STIR; 19:49 min) and high-resolution DL-DIXON imaging (1 mm isotropic; 5:24 min). Three blinded readers assessed overall imaging diagnosis, diagnostic confidence, and lesion presence (osteitis, fat metaplasia, erosions, sclerosis, joint space changes). DL was used solely for image reconstruction. The clinical diagnosis served as the reference for diagnostic accuracy; sMRI served as the reference for lesion analysis. Diagnostic accuracy was expressed as AUC (balanced accuracy). Non-inferiority was tested using a predefined margin of 0.05. A decrease of ≤0.05 compared to sMRI was defined as clinically acceptable. Nineteen patients were diagnosed with axSpA. DL-DIXON showed similar diagnostic performance (AUC 0.86 [95% CI 0.76-0.96]) compared to sMRI (0.87 [0.78-0.96]) and met non-inferiority criteria. Interreader agreement was almost perfect for sMRI (κ = 0.81) and substantial for DL-DIXON (κ = 0.71). At the lesion level, DL-DIXON achieved good performance for erosions (AUC 0.83), fat metaplasia (0.80), and joint space changes (0.80), and fair performance for osteitis (0.73) and sclerosis (0.64). Using DL-DIXON, scan time was reduced by 73%. DL-DIXON provides clinically acceptable diagnostic performance while substantially reducing scan time. This approach may improve efficiency and accessibility of MRI for axSpA assessment.