Deep learning based CT grading system for sacroiliitis: a multi-center studydemonstrating superior accuracy and efficiency compared to human readers.

BACKGROUND

To develop and validate a deep convolutional neural network (DCNN) for automated sacroiliitis grading in axial spondyloarthritis (axSpA) using CT images.

METHODS

A total of 1,590 patients were enrolled, including 1341 axSpA patients from a Rheumatology Specialist Hospital for 3D-ResNet50 model development. The model was then evaluated on an internal validation set of 130 patients and external sets from two tertiary hospitals (n = 249). Diagnostic ssensitivity and reading time were assessed for six readers (two rheumatologists, two junior radiologists, and two senior radiologists). Participant-assigned sacroiliitis Grades were compared with model predictions, and the area under the curve (AUC) was analyzed using the DeLong test.

RESULTS

The 3D-ResNet50 model achieved overall sample-level diagnostic accuracies of 89.6 %, 89.9 %, 88.5 %, and 87.1 % in the training, testing, internal validation, and external validation sets, respectively.The AUCs for Grade I, II, III, and IV in external validation set were 0.93, 0.89, 0.92 and 0.88, all exceeding those of readers (AUC: 0.65-0.93), with the most significant improvement at Grade II (p < 0.001 for Grades I-III; p > 0.05 for Grade IV). The model significantly reduced diagnostic time per case (2.74 ± 0.15 s) compared to readers (119.4 ± 42.0 s).

CONCLUSION

The 3D-ResNet50 model accurately grades sacroiliitis with superior sensitivity and efficiency compared to rheumatologists and radiologists. Its strong performance supports its integration into clinical practice to assist in axSpA diagnosis.