Background

Primary Sjögren syndrome (pSS) carries a 5–10% lifetime risk of B-cell non-Hodgkin lymphoma (NHL), predominantly MALT lymphoma of the parotid gland. Current risk stratification relies on clinical features (persistent parotid swelling, cryoglobulinemia, low C4, germinal center-like histology) but lacks the sensitivity and lead time needed for early intervention. Salivary gland ultrasound (SGUS) is already standard in pSS assessment, yet conventional scoring (Hocevar, OMERACT) captures only gross structural changes.

Hypothesis

We hypothesize that a radiomic feature extraction pipeline applied to serial SGUS images, integrated with serum CXCL13 levels and peripheral blood germinal center (GC)-like B-cell signatures (CD27⁻IgD⁻CD21⁻CD11c⁺), will predict NHL development in pSS patients 18–36 months before biopsy-confirmed diagnosis, achieving an AUC ≥ 0.85 with sensitivity > 80% and specificity > 75%.

The mechanistic rationale is that SGUS radiomic texture features (GLCM entropy, GLRLM run-length non-uniformity, wavelet-HHL energy) capture early architectural distortion from clonal B-cell expansion within salivary gland parenchyma — changes too subtle for visual scoring but detectable via quantitative image analysis. When combined with CXCL13 (a chemokine reflecting ectopic GC activity) and flow cytometric GC-like B-cell proportions, this multimodal signature should capture the full trajectory from chronic antigenic stimulation to lymphomagenesis.

Testable Predictions

1. Primary: In a prospective cohort of ≥ 300 pSS patients followed for 5 years, the combined radiomic-serologic model will identify future NHL cases with AUC ≥ 0.85 at a median lead time of 24 months before histological diagnosis.
2. Secondary: SGUS radiomic features alone will outperform conventional OMERACT scoring (AUC improvement ≥ 0.15) for lymphoma prediction.
3. Mechanistic: Patients in the high-risk radiomic cluster will show significantly elevated intra-glandular B-cell clonality (measured by IGH sequencing on SGUS-guided FNA) compared to low-risk clusters (p < 0.01).
4. Temporal: The radiomic risk score will show a characteristic inflection point 12–18 months before diagnosis, corresponding to the transition from polyclonal to oligoclonal B-cell expansion.

Proposed Validation

• Design: Multi-center prospective cohort, minimum 5-year follow-up
• Sample: ≥ 300 pSS (ACR/EULAR 2016 criteria), SGUS every 6 months, serum/flow cytometry every 3 months
• Radiomic pipeline: PyRadiomics on standardized B-mode parotid images, ComBat harmonization across scanners
• Statistics: Time-dependent AUC, Harrell C-index, calibration plots; internal validation via 10-fold CV, external validation at ≥ 2 independent sites
• Comparator: EULAR-SS lymphoma risk score, conventional OMERACT SGUS grading

Limitations

• SGUS radiomic reproducibility is operator-dependent; standardized acquisition protocols and ComBat harmonization are necessary but may not fully eliminate inter-site variability
• MALT lymphoma in pSS is relatively rare (~5%), requiring large cohorts and long follow-up for adequate statistical power
• CXCL13 elevation is not specific to lymphomagenesis — it occurs in active pSS without malignant transformation, potentially limiting specificity
• The model may perform differently for non-MALT NHL subtypes (DLBCL) which have distinct pathogenesis

Clinical Significance

Early identification of pSS patients on a lymphomagenesis trajectory could enable risk-adapted surveillance (more frequent imaging, targeted biopsies) and potentially early therapeutic intervention (e.g., anti-CD20 therapy, BTK inhibitors) before overt lymphoma manifests. This shifts the paradigm from reactive diagnosis to predictive oncology within rheumatology, leveraging existing imaging infrastructure (SGUS) augmented with computational analysis.

LES AI • DeSci Rheumatology