Document resource
Background Endoscopic ultrasound-guided fine-needle aspiration/biopsy (EUS-FNA/B) is essential for diagnosing pancreatic solid lesions (SPLs), but rapid on-site evaluation (ROSE) is often limited by the shortage of cytopathologists and the demands of real-time assessment. We developed and validated a whole-slide imaging (WSI)-based artificial intelligence (AI) system for automated cytologic classification of Diff-Quik-stained EUS-FNA/B smears ( IDDF2026-ABS-0380 Figure 1. The digital pathology scanner and AI diagnostic system of the study).Methods This study included a retrospective pilot phase followed by single-center model development and multicenter external validation ( IDDF2026-ABS-0380 Figure 2. Study design and AI workflow). The pilot phase included 312 pathologically confirmed SPL cases from Ruijin Hospital for model pretraining and methodological validation. The development cohort comprised 870 WSIs from 327 patients, including the pilot cases. External validation used 256 WSIs from 98 patients across six independent centers. Final diagnoses were established by surgical pathology, core biopsy histology, or clinical follow-up. The proposed two-stage multi-scale network first localized diagnostically relevant cell clusters at 5× magnification and then performed three-class classification at 20× using multiple-instance learning (IDDF2026-ABS-0380 Figure 3. Two-stage AI workflow for cytologic classification of SPLs). The target categories were pancreatic ductal adenocarcinoma (PDAC), pancreatic neuroendocrine tumor/carcinoma (PNET/C), and benign lesions. The primary endpoint was overall accuracy; secondary endpoints included sensitivity, specificity, F1 score, and AUC.Results In the pilot cohort, overall accuracy and AUC were 0.930 and 0.948, respectively. In the expanded Ruijin cohort, internal split testing achieved slide-level accuracy of 0.885 and AUC of 0.902; patient-level accuracy was 0.849. In the independent external multicenter cohort, the model achieved a slide-level overall accuracy of 0.855 and AUC of 0.915; the patient-level overall accuracy was 0.878. ( IDDF2026-ABS-0380 Figure 4. Diagnostic performance of the AI system in internal and external validation cohorts). Performance was strongest for PDAC and benign lesions, whereas classification of PNET/C was comparatively lower. High-risk cell-cluster visualization improved model interpretability (Figure 5. Representative visualization of high-risk cell clusters).Conclusions This WSI-based two-stage AI system showed accurate cytologic classification of SPLs and good multicenter generalizability. It may serve as an interpretable adjunct to ROSE.Abstract IDDF2026-ABS-0380 Figure 1Abstract IDDF2026-ABS-0380 Figure 2Abstract IDDF2026-ABS-0380 Figure 3Abstract IDDF2026-ABS-0380 Figure 4Abstract IDDF2026-ABS-0380 Figure 5