Detective flow imaging (DFI) is an emerging endoscopic ultrasound (EUS) modality capable of visualizing low-velocity microvascular flow without contrast agents (1). While qualitative assessment of pancreatic vascularity has been explored (1-4), quantitative metrics such as the number of vessels are insufficient. Automated vessel-detection using computer-assisted detection (CADe) software may allow objective measurement of microvascular density, potentially distinguishing healthy pancreas from inflammatory or neoplastic pathology. We aim to evaluate the diagnostic accuracy of a CADe developed for quantification of fine vessels within the pancreas during real-time EUS.

A prospective study including consecutive adult patients undergoing EUS was performed between June 2024 and October 2025. The pancreas was assessed using DFI. Real-time CADe software analyzed live DFI images, detecting and counting fine vessels. Average fine vessel count (AVC) was defined as the mean vessel count detected by the CADe during the procedure. Ten external physicians, divided into senior (40%) and fellow (60%) endoscopists, evaluated 11 randomly selected EUS-DFI videos to provide vessel counts and categorical interpretation. Intra-class correlation (ICC) and interobserver agreement (IOA) for interpretation were calculated.

Sixty-five participants were included (Inflammatory: 35; Neoplastic: 17; Healthy: 13). Mean age was 53.5±15.0, 66.2±15.9 and 42.2 ± 16.1 (p<.05), and 58.5% were female (p=.281). There were no differences between abdominal pain, and nausea/vomiting among groups (p>.05), jaundice was more common in the neoplastic lesions (p=.03) (Table 1).

CADe provided a median AVC of 13.0 (10-17) for inflammatory conditions, 8.0 (7.0-16.0) for neoplastic, and 4.0 (3.0-6.0) for healthy pancreas (p<0.05) (Figure 1). The maximum fine vessel count (MVC) was 36.0 (28.0-48.0) for inflammatory conditions, 33.0 (21-36) for neoplastic, and 14.0 (11-23) for healthy pancreas (p<0.05). Sensitivity, specificity, positive and negative predictive values, and observed agreement for distinguishing between pathological from healthy pancreas were 90%, 92%, 98%, 71%, and 91%, respectively. Visual assessment showed poor ICC in both senior (-0.2, p=.965) and fellow (-0.09, p=.905), whereas categorical interpretation achieved substantial agreement (senior: k=0.624, p<0.05; fellows: 0.653, p<0.05). 

CADe-based vessel detection provides reliable, objective quantification of fine pancreatic vessels during real-time EUS, demonstrating high diagnostic accuracy for differentiating pathological from healthy pancreas. Inflammatory lesions exhibited higher AVC than neoplastic lesions, which had values closer to healthy tissue, suggesting reduced microvascularity in pancreatic neoplastic lesions. Visual assessment showed poor reproducibility for vessel counts but substantial agreement for categorical interpretation.