Artificial intelligence improves accuracy of pancreatic screening EUS by enhancing the detection of small solid lesions: a tandem randomized video study

Aims To evaluate the add-on effect of a novel Artificial Intelligence (AI) software system (mAI Companion©, Medi-Globe GmbH, Germany) on the performance of endosonographers in detecting small solid pancreatic tumours (SPTs) during endoscopic ultrasound (EUS) examinations. Methods A randomized, controlled tandem video study was conducted. Sixteen EUS videoclips were selected, each containing a standardized complete pancreatic examination of small SPTs (9-19mm) in different locations of the pancreas. Each clip included case-related contextual information, followed by a frozen frame at the end, with or without a SPT. Each video was assessed twice by every participant: once with AI assistance and once without. Randomization (1:1) determined whether the first or second evaluation included AI assistance. For each clip, participants were asked to assess the frozen frame and to mark the perceived lesion localization on a paper grid. Their assessments were compared with the ground truth established by an expert endosonographer. Results Thirty-two endosonographers with varying levels of experience in EUS participated. Stand-alone AI showed a diagnostic accuracy of 93.75%. Without AI support, participants achieved an overall accuracy of 70.51% (95%CI 66.35%-74.43%). With AI assistance, performance significantly improved to 82.03% (95%CI 78.43%-85.26%). The accuracy gain with AI was statistically significant (p