A SOFTWARE SYSTEM FOR BUILDING GASTROINTESTINAL ENDOSCOPY DATASETS: CLINICAL ANNOTATION SCHEMA AND AUTOMATED QUALITY CONTROL

Abstract. The quality of deep learning models for gastrointestinal endoscopy depends on how the training dataset is assembled, labelled and split no less than on model architecture. In practice, groups combine a general-purpose annotation tool with ad-hoc spreadsheets, which leads to duplicated patients, inconsistent lesion descriptions, leakage of images from one patient between training and test subsets, and annotation noise caused by optical artifacts. We present a desktop workflow that couples a local patient and image registry with the CVAT annotation platform. It (i) keeps identifying data on the local machine and exports only de-identified metadata; (ii) encodes a clinical anatomical and lesion taxonomy as a CVAT label schema; (iii) introduces a single “artifact” class so that models learn to ignore non-diagnostic regions; (iv) imports COCO annotations into relational tables, keeping bounding boxes and polygon masks from one annotation pass; and (v) runs automated quality control, including a patient-leakage check, before export to YOLO (detection and segmentation) and COCO formats. The design is described with IDEF0, IDEF1X and IDEF3 models and compared with existing tools and public datasets. Keywords: endoscopy; medical image dataset; data curation; annotation workflow; CVAT; data leakage; object detection; instance segmentation.