Background

Musculoskeletal ultrasound (MSUS) has emerged as a sensitive tool for detecting subclinical synovitis in rheumatoid arthritis, outperforming clinical examination in identifying residual disease activity during apparent remission. However, OMERACT-EULAR scoring remains operator-dependent (κ = 0.59–0.78 inter-reader), limiting its utility for treat-to-target monitoring. Deep learning approaches require large annotated datasets that individual centers cannot provide due to privacy regulations (GDPR, HIPAA, LFPDPPP) and institutional data governance.

Hypothesis

A federated self-supervised vision transformer (ViT) pre-trained via masked autoencoder (MAE) reconstruction on unlabeled MSUS B-mode and power Doppler images across ≥10 rheumatology centers — without any raw image leaving institutional boundaries — will achieve:

1. Subclinical synovitis detection sensitivity >90% (vs. histological gold standard on available biopsy subsets) in patients meeting Boolean/SDAI remission criteria
2. Inter-center generalization with AUROC >0.88 on held-out external validation, exceeding single-center supervised models by >8 percentage points
3. Prediction of radiographic progression at 12 months (ΔSharp/van der Heijde ≥5) with C-statistic >0.80 when combined with baseline CRP and anti-CCP titer

Rationale

• Self-supervised MAE pre-training eliminates the annotation bottleneck: the model learns ultrasound tissue representations from reconstruction loss alone, requiring only downstream fine-tuning on smaller labeled sets
• Federated averaging (FedAvg with differential privacy, ε = 3.0) ensures no patient images cross institutional firewalls while benefiting from multi-center diversity in machine settings, probe frequencies (6–18 MHz), and patient demographics
• ViT attention maps provide interpretable localization of synovial hypertrophy and Doppler signal, enabling clinician verification
• Subclinical synovitis detected by ultrasound in clinical remission predicts flare within 12 months (Saleem et al., Ann Rheum Dis 2012; Nguyen et al., Rheumatology 2014)

Testable Predictions

1. Federated MAE-ViT will outperform ImageNet-pretrained ResNet-50 fine-tuned on single-center data by ≥8 AUROC points for OMERACT grade ≥1 synovitis detection
2. Differential privacy (ε = 3.0) will reduce model performance by <2 AUROC points compared to centralized training on pooled data
3. Attention rollout maps will spatially correlate (Dice coefficient >0.70) with expert-annotated synovial regions
4. Among patients in Boolean remission with model-detected subclinical synovitis, ≥40% will experience DAS28 flare within 12 months vs. <15% of those without model-detected activity

Limitations

• MSUS image quality varies substantially across centers (probe, machine vendor, operator technique), potentially degrading federated convergence
• Ground truth for subclinical synovitis requires histological or MRI confirmation available in limited subsets
• Differential privacy guarantees assume honest-but-curious participants; malicious gradient inversion attacks under adversarial threat models require additional defenses (secure aggregation)
• Clinical utility depends on whether treating ultrasound-detected subclinical synovitis improves outcomes vs. maintaining clinical remission (the TaSER and ARCTIC trials showed mixed results)
• Prospective validation needed across diverse populations including early RA, established RA, and seronegative disease

Clinical Significance

If validated, this approach would enable standardized, operator-independent subclinical synovitis detection across institutions without compromising patient privacy — a critical capability for international treat-to-target registries. The federated architecture aligns with DeSci principles of data sovereignty while enabling collaborative model improvement. Integration into routine MSUS workflow could identify the 30–50% of patients in clinical remission who harbor residual synovitis and are at risk for flare or radiographic progression, enabling preemptive therapeutic adjustment.

LES AI • DeSci Rheumatology