Anatomy-centred deep learning improves generalisability and progression prediction in radiographic sacroiliitis...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Axial spondyloarthritis (axSpA) diagnosis relies on radiographic assessment.
• Improving the generalisability and predictive power of deep learning models for axSpA is crucial.

Purpose of the Study:

• To evaluate if anatomy-centred deep learning enhances model generalisability for detecting radiographic sacroiliitis.
• To determine if this approach improves prediction of disease progression.

Main Methods:

• Retrospective multicentre study using pelvic radiographs from four patient cohorts.
• Trained two deep learning models: one anatomy-centred (sacroiliac joints) and one standard (full radiograph).
• Compared model performance using AUC, accuracy, sensitivity, and specificity on independent test datasets; assessed progression prediction using follow-up data.

Main Results:

• The anatomy-centred model demonstrated superior performance across all test datasets (AUCs: 0.899-0.957 vs. 0.853-0.947).
• Anatomy-centred model achieved higher accuracy (0.821-0.906 vs. 0.770-0.850).
• High-risk patients identified by the anatomy-centred model had a 2.16-fold increased odds of radiographic sacroiliitis progression within 2 years.

Conclusions:

• Anatomy-centred deep learning significantly improves the generalisability of models for detecting radiographic sacroiliitis.
• This method enhances prediction of disease progression in axial spondyloarthritis.
• The developed model is available as open source.