The Dresden Endoscopy Dataset for Endoscopic Retrograde Cholangiopancreatography

Aims Artificial intelligence (AI) is rapidly advancing diagnostics and therapeutic capabilities in medical imaging. In gastrointestinal endoscopy, AI shows promise in polyp and cancer detection, particularly in high-frequency procedures like colonoscopy or esophagogastroduodenoscopy. However, for advanced interventions like endoscopic retrograde cholangiopancreatography (ERCP), progress is limited by the lack of large, high-quality datasets. To address this gap, we introduce the Dresden Endoscopy Dataset for ERCP, comprising 1,000 ERCP recordings with video-wise annotations. This dataset aims to provide a valuable resource for the scientific community to foster AI model development, ultimately assisting endoscopists and enhancing patient safety in ERCP.

Methods Recording devices were installed in the fluoroscopy suites at the University Hospital Carl Gustav Carus in Dresden, Germany, to capture video signals from the endoscopic processor. ERCP procedures were recorded from routine clinical practice starting in January 2021. Recordings underwent semi-automated processing to yield cropped, trimmed videos, removing extraluminal footage and blurring nuisance text from the video processor. Each video was paired with the examination report and demographic information within a research data management platform. Unstructured reports were converted into a machine-readable format using structured forms completed by trained student assistants and validated by two gastroenterologists. A multiple instance learning (MIL) model was trained and employed on a subset of videos to validate the dataset for an ERCP-specific AI task.

Results From January 2021 to September 2023, videos of 1,023 ERCP procedures were recorded. Semi-automated preprocessing of videos, combined with variables extracted from the written reports, yielded a high-quality dataset suitable for developing and testing ERCP specific AI models. The structured data extracted from the reports included key procedural elements observable in the endoscopy video, such as: stent-placement, significant bleeding, discharge from the papilla, anatomical variations, papillotomy, and use of various endoscopic tools and devices. As one possible application example, the dataset was validated by employing the MIL model on a stent-detection task. The model achieved a precision of 80%, specificity of 79%, an overall accuracy of 79%, and an area under the receiver operating characteristic curve (AUROC) of 80% in detecting stent presence.

Conclusions The proposed dataset represents the largest collection of ERCP procedures with video-wise annotations, providing a valuable resource for AI model development. Preliminary technical validation confirms the dataset’s usability for training clinically relevant AI models, addressing a critical gap in the limited landscape of high-quality, publicly available ERCP datasets. Future datasets could be enriched by incorporating synchronized fluoroscopy video streams, which may provide even greater insights, enhancing the potential for robust and clinically valuable AI applications in complex endoscopic procedures.

Publication History

Article published online:
27 March 2025