Generating synthetic endoscopy videos following a systematic screening protocol

Gastric diseases are a considerable burden world-wide and Upper endoscopy or Esophagogastroduodenoscopy (EGD) is the gold standard for their diagnosis, a procedure highly dependent on the gastroenterologist’s experience. Overall, 11.3% of lesions are misdiagnosed during a routine EGD, despite a systematic stomach screening protocol (SSS) is followed to avoid blind spots. EGD diagnosis may be supported by providing annotated video collections useful to develop and compare computational strategies as second readers for SSS, or as study cases for trainees. However, generating videos with SSS annotations is a highly time-consuming task for an expert, and none of the existent collections are publicly available. Therefore, virtually modeling the EGD procedure may allow the generation of synthetic videos with SSS annotations. In this work, a collection of synthetic videos simulating EGD procedures is constructed in a virtual environment. A stomach model is developed upon reported averaged measurements and modified to achieve different realistic appearances by integrating organ deformations, gastric rugae, inner stomach irregularities, and tissue appearance. A virtual camera and light source are displaced within the modified stomach model following a path that emulates a typical SSS, generating an annotated video with 22 anatomic SSS landmarks. Likewise, shape information, i.e., a depth map for each frame, is provided. Finally, 50 videos were rendered with 315,000 frames and released for public access, available at http://cimalab.unal.edu.co/?lang=en&mod=software.