- Design
- two-stage deep learning model development with external validation on two independent phase 3 trial datasets, no retraining
- Population
- trained on 132 patients (MEASURE 1); tested on 555 (PREVENT) and 414 (SURPASS) with axial spondyloarthritis
- Primary outcome
- detection of five MRI-defined sacroiliac lesions against expert Berlin-method reading
- Effect
- ankylosis AUC 0.97-0.99 with balanced accuracy 0.95-0.97; bone marrow oedema AUC 0.85-0.93 with balanced accuracy 0.72-0.74; overall comparable with expert interreader agreement
Scoring sacroiliac joint MRI in axial spondyloarthritis is slow and notoriously reader-dependent, which is a problem for trials before it is a problem for clinics. A two-stage deep learning pipeline was built to outline both sacroiliac joints on paired T1-weighted and STIR sequences and then detect five lesion types: bone marrow oedema, and the structural lesions erosion, fat lesion, sclerosis and ankylosis.
It was trained on 132 patients from the MEASURE 1 trial using consensus labels, then tested without further training on two entirely separate trial datasets — PREVENT (555 patients) and SURPASS (414). Against expert reading by the Berlin method, performance was comparable with the agreement experts reach with each other. Structural lesions did best, ankylosis especially (AUC 0.97 and 0.99, balanced accuracy 0.95 and 0.97). Bone marrow oedema had consistently high AUCs of 0.85 to 0.93 but balanced accuracy of only 0.72 to 0.74.
That gap between AUC and balanced accuracy for oedema is the honest limitation, and it matters because oedema is the lesion that drives treatment decisions. The system ranks scans well but does not classify individual joints as confidently, and it was validated in trial populations with protocol-standardised imaging, not on a district hospital's scanner. This is a tool for making trial and cohort scoring consistent and scalable. It is not yet something to put between a radiologist and a patient's report.
- Read this as trial infrastructure, not as a diagnostic aid for clinical reporting
- Structural lesions are read most reliably; bone marrow oedema least — and oedema is what changes treatment
- Generalisation held across two independent trial datasets without retraining, which is the strongest part of the result
- Protocol-standardised trial imaging is not the same as routine scans; local validation would be needed
- Keep MRI interpretation anchored to the clinical picture: lesions on imaging do not by themselves define axSpA
The statistics, in plain English
An AUC of 0.97 for ankylosis means the model almost always ranks an affected joint above an unaffected one. Balanced accuracy is the stricter measure — it asks how often the model is right after a threshold is chosen, averaged over positive and negative joints — and for bone marrow oedema it sat at 0.72 to 0.74 despite an AUC of 0.85 to 0.93. That combination indicates a discrimination problem at the decision threshold rather than in the ranking. Benchmarking against expert interreader agreement rather than a true reference standard means the ceiling is human variability, not truth.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for rheumatology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free