Sjögren's syndrome-focusassist: lymphocytic focus assessment in Sjögren's syndrome: a deep learning and spatial analysis approach.

Sjögren’s syndrome (SS) is a chronic autoimmune disease whose clinical gold standard relies on precise quantification of lymphocyte foci and lymphocyte density in salivary gland biopsies. However, traditional manual assessment is time-consuming, highly subjective, and lacks consistency. To address these challenges, we propose Sjögren’s syndrome-FocusAssist (SS-FocusAssist), an end-to-end AI-assisted pathology framework for automated focus detection and quantification. At the cell-level detection stage, the system adopts an anchor-free single-stage detector enhanced with Coordinate Attention, improving the model’s ability to capture long-range spatial dependencies and identify minute lymphocytes. This design yields a precision of 95.9% and an mAP@0.5 of 99.1%, representing improvements of 11.1% and 7.3% over the baseline model. At the lesion-level aggregation stage, we introduce Filter-guided Adaptive Clustering (FAC): KDTree is first used to suppress spatial outliers, followed by DBSCAN with adaptive thresholds to form density-consistent lymphocytic foci; Finally, in the quantification stage, we develop an iterative expert-algorithm threshold optimization framework, guided by clinical diagnostic criteria, to achieve an optimal balance between lesion detection fidelity and pathological interpretability. In an independent cohort of 298 cases, SS-FocusAssist achieved a Cohen’s κ of 0.793 compared with dual-expert annotations—a 39% improvement over cell-detection-only methods. These results demonstrate that SS-FocusAssist substantially enhances the accuracy, consistency, and efficiency of SS biopsy interpretation, offering a clinically practical pathway for scalable digital pathology deployment.