Pancreatic ductal adenocarcinoma (PDAC) is difficult to detect early, yet the chance of curative treatment depends heavily on finding disease while it is still resectable. Small lesions can be missed on conventional imaging because the pancreas lies deep in the abdomen and is difficult to visualise consistently. Endoscopic ultrasound (EUS) provides high-resolution imaging close to the pancreas and can detect very small lesions, making it an important tool for diagnosis and staging.
EUS is also highly operator-dependent. The endoscopist must obtain and interpret multiple partial views while moving through different pancreatic stations, and diagnostic performance varies with experience. This matters particularly when PDAC resembles inflammatory masses associated with chronic pancreatitis. A missed or uncertain lesion may delay diagnosis beyond the narrow window in which surgery remains possible, while interpretation based on a few selected still frames does not reflect how an endoscopist reaches a patient-level judgement across the examination.
Endoscopists need support that works during the examination, using the routine B-mode EUS stream without requiring them to stop and select a lesion or acquire an additional imaging mode. The system should recognise when the pancreas is actually being examined, combine information across the different views obtained during the procedure, and return one patient-level assessment for the examination. It also needs to operate with sufficiently low latency to be useful while the scope is still in the patient and to function as decision support rather than replacing clinical judgement.
We are developing an automated system that reads routine B-mode EUS video and returns one patient-level result for the whole examination, fast enough to be used while the scope is still in the patient. Development uses a retrospective single-centre cohort of 279 patients spanning confirmed PDAC, benign inflammatory masses, and patients without a visible tumour, evaluated entirely at the patient level.
In the primary tumour cohort (n=97), the complete gated pipeline achieved 85.57% accuracy, 92.45% sensitivity, 77.27% specificity, and an ROC-AUC of 84.05%. Across all 279 patients, ROC-AUC was 93.41% and NPV was 98.05%. Inference latency was below 50 ms on a laptop-class GPU, and sensitivity reached 84.90% after 60% of the procedure. A separate 12-case pilot reader study supported feasibility but was not powered to establish superiority over endoscopists.