This study aimed to develop and validate a deep learning-based system for automated quantification of lymphocytic foci and FS calculation in LSGB specimens.

This study enrolled patients with ocular and/or oral dryness meeting AECG criteria who underwent LSGB at Sun Yat-Sen Memorial Hospital, with ethical approval and ClinicalTrials.gov registration (NCT06437652). Three senior pathologists established ground truth through blinded consensus review using QuPath software (v0.4.4) to annotate salivary gland areas and quantify lymphocytic foci (≥50 cells/focus) in whole-slide images. The LabialFS system was developed with three deep learning algorithms for gland segmentation, lymphocytic foci identification, and nuclei counting, trained on H&E-stained whole-slide images (WSIs) scanned at 20×resolution (0.5 μm/pixel) (Figure 1). The operational workflow involved glandular tissue segmentation, lymphocytic foci detection within demarcated areas, nuclei verification, and automated focus score calculation (FS = foci count/gland area × 4). Internal validation used 200 consecutive specimens, while external validation employed 484 WSIs from three medical centers with different scanning platforms (KF-PRO-005-EX, KF-PRO-120, Aperio CS2). The system was deployed as a second-read quality control tool with 90% specificity threshold, with performance metrics (AUC, sensitivity, specificity, accuracy) calculated against ground truth using MedCalc and SAS software.

During the development and internal testing phase, the LabialFS system successfully integrated three deep learning models—gland segmentation, lymphocytic foci segmentation, and nuclei segmentation—trained on 119 WSIs with 2,037 mm² of annotated salivary gland tissue and 167 WSIs containing 2,078 labeled lymphocytic foci. The gland segmentation model achieved a Dice score of 0.953±0.054, while the lymphocytic foci segmentation model demonstrated robust performance with a Dice score of 0.729±0.081 and an F1-score of 0.790±0.093. In internal validation using 200 consecutive WSIs, the system exhibited exceptional performance in FS quantification, achieving an AUC of 0.98 (95% CI: 0.961–0.999) for determining FS≥1, with 98% accuracy, 99% specificity, 97% sensitivity, 99% PPV, and 97% NPV, significantly outperforming routine histopathological reports which showed only 71% accuracy. The system effectively addressed common limitations of manual assessment, such as overestimation of FS in 9.5% of cases due to misinterpretation of ductal structures and underestimation in 3% of cases caused by difficulties in counting lymphocytes near the 50-cell threshold, while showing strong correlation with ground truth FS values (R²=0.9409, p<0.001). External validation across 484 WSIs from three independent medical centers confirmed the system's robustness, with an overall AUC of 0.941 (95% CI: 0.920–0.962), 94.2% accuracy, 99.6% sensitivity, and 88.7% specificity, and center-specific analyses revealed consistent performance across different scanner platforms and staining protocols. In clinical deployment as a second-read quality control application, the LabialFS system provided significant value by detecting lightly stained lymphocytic foci missed by junior pathologists, accurately differentiating epithelial structures from true lymphocytic aggregates, and maintaining high performance across diverse lymphocytic focus morphologies, thereby addressing the critical need for standardized FS assessment protocols in settings with limited specialist expertise in autoimmune pathology.

The LabialFS system is a rigorously validated AI tool that automates FS calculation with high precision, addressing critical limitations of manual pathology assessment. By standardizing SjD diagnosis across diverse clinical settings, it enhances early detection and reduces inter-observer variability. The system’s integration into routine pathology workflows supports scalable, accurate SjD diagnostics, with potential to improve patient outcomes through timely intervention.