Spondyloarthritis & psoriatic arthritis
Anatomy-centred deep learning improves generalisability and progression prediction in radiographic sacroiliitis detection
Authors' conclusion
Anatomy-centred deep learning can improve the generalisability of models in detecting radiographic sacroiliitis.
Study context
- Population
- Conventional pelvic radiographs of four patient cohorts focusing on axSpA, university and community hospitals4 cohorts
- Comparison
- Anatomy-centred deep learning model vs standard model
- Outcome
- Detection of radiographic sacroiliitis: AUC, accuracy, sensitivity and specificity on three independent test datasets
Methodological detailsretrospective, multicentre, Model published as fully open source, First cohort of 1483 radiographs split into training (n=1261) and validation (n=222) sets; other cohorts of 436, 340 and 163 patients used as independent test datasets
- Retrospective
- Multicentre
- Model published as fully open source
- First cohort of 1483 radiographs split into training (n=1261) and validation (n=222) sets; other cohorts of 436, 340 and 163 patients used as independent test datasets
Performance
As reported in the abstract
Radiographic sacroiliitis detection on the three test datasets (respectively)
| Test or model | AUC (test datasets 1, 2, 3) | Accuracy (test datasets 1, 2, 3) |
|---|---|---|
| Anatomy-centred | 0.899, 0.846, 0.957 | 0.821, 0.744, 0.906 |
| Standard | 0.853, 0.817, 0.947 | 0.770, 0.724, 0.850 |
AUC (test datasets 1, 2, 3)
Anatomy-centred0.899, 0.846, 0.957
Standard0.853, 0.817, 0.947
Accuracy (test datasets 1, 2, 3)
Anatomy-centred0.821, 0.744, 0.906
Standard0.770, 0.724, 0.850
Further findings
Progression of radiographic sacroiliitis, high risk vs not high risk by the anatomy-centred model (second cohort follow-up, 311 patients) · 2 years
| OR (95% CI) | ||
|---|---|---|
| High risk by anatomy-centred model | 2.16 (1.19–3.86) | |
| 0.20.5125Odds ratio (log scale) |
Original publication
Anatomy-centred deep learning improves generalisability and progression prediction in radiographic sacroiliitis detection. RMD Open. Published 23 December 2024. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →