AI-Assisted Real-Time Cytologic Diagnosis During EUS-FNA of Pancreatic Masses (With Video)

Rapid on-site evaluation enhances the diagnostic yield of endoscopic ultrasound-guided fine needle aspiration (EUS-FNA), but the cytopathologists are limited. This study aimed to assess the diagnostic capability of an artificial intelligence-assisted ROSE (AI-ROSE) for EUS-FNA.

The study included 137 patients who underwent EUS-FNA of a pancreatic mass between April 2019 and August 2021. Participants were divided into training (n = 96), validation (n = 15), and test cohorts (n = 26). From the training/validation cohort, 5157/615 digital images of cell clusters were extracted and divided into 288 × 288-pixel patches. These cell clusters were annotated and labeled into five classes. A semantic segmentation architecture was developed. From the test cohort, 120 cell clusters were extracted to compare diagnostic performance between AI-ROSE and 21 endosonographers and 5 cytotechnologists with varying experience levels in ROSE.

In total, 1,097,840 training, 31,817 validation, and 1920 test regions were extracted. In the test cohort, AI-ROSE accuracy for three-category classification (class 1/2, class 3, and class 4/5) was 89.8%. For two-category classification (class 1/2/3 and class 4/5), sensitivity, specificity, and accuracy were 89.3%, 98.1%, and 95.1%, respectively. In the comparison cohort, AI-ROSE accuracy for two-category classification was 93.3%, significantly higher than all endosonographers (68.3%; range, 45.8%-86.7%) and cytotechnologists (76.3%; range, 72.5%-78.3%). The AI-ROSE evaluation time for 120 cell clusters was 6.04 s, much shorter than that of all endosonographers (1800; 480-6000 s) and cytotechnologists (2160; 1020-3600 s).

The AI-ROSE model shows remarkable speed and accuracy in diagnosing pancreatic cell clusters, enabling rapid decision-making during EUS-FNA. UMIN-CTR; No. 000042212.