Deep learning-based synthetic CT for the assessment of spinal structural lesions in axial spondyloarthritis.

Abstract

Objective: To evaluate the accuracy of synthetic CT for detecting spinal structural lesions in patients with suspected axial spondyloarthritis (axSpA).

Materials and methods: Patients with inflammatory back pain who underwent both spinal CT and MRI examinations were included in this retrospective study. Two readers independently assessed the presence of structural lesions (facet joint ankylosis, structural Andersson lesions, discal ossifications, vertebral ankylosis, syndesmophytes, and erosions) on conventional CT and synthetic CT images. Feature detection accuracy of synthetic CT relative to conventional CT was evaluated using sensitivity, specificity, positive predictive value, and negative predictive value, with consensus reading as the reference standard. In addition, lesion-level erosion detection accuracy was assessed, and erosion size was compared between conventional CT and synthetic CT.

Results: Eighty-seven patients (mean age, 41 ± 14.2 years; range, 18-80 years; 42 males and 45 females) were included. Sensitivity of synthetic CT was 1.00 for facet joint ankylosis, structural Andersson lesions, vertebral ankylosis, and erosions, 0.91 for syndesmophytes, and 0.67 for discal ossifications. Specificity ranged from 0.97 to 1.00 across all lesion types. Lesion-level analysis showed sensitivity of 0.82-1.00 and specificity of 0.96-1.00 for erosion detection across different spinal regions. No significant differences of erosion size (median [IQR], mm) between conventional CT and synthetic CT (width, 7.7 [4.4-9.7] vs. 8.0 [4.1-9.6], P = 0.15; depth, 4.5 [3.5-6.7] vs. 4.5 [3.5-6.5], P = 0.12).

Conclusion: Synthetic CT demonstrates a high feature detection accuracy for assessing most spinal structural lesions in patients with suspected axSpA.