Abstract
BACKGROUND AND AIMS: Capsule endoscopy (CE) has revolutionized small-bowel (SB) investigation. Computational methods can enhance diagnostic yield (DY); however, incorporating machine learning algorithms (MLAs) into CE reading is difficult as large amounts of image annotations are required for training. Current databases lack graphic annotations of pathologies and cannot be used. A novel database, KID, aims to provide a reference for research and development of medical decision support systems (MDSS) for CE. METHODS: Open-source software was used for the KID database. Clinicians contribute anonymized, annotated CE images and videos. Graphic annotations are supported by an open-access annotation tool (Ratsnake). We detail an experiment based on the KID database, examining differences in SB lesion measurement between human readers and a MLA. The Jaccard Index (JI) was used to evaluate similarity between annotations by the MLA and human readers. RESULTS: The MLA performed best in measuring lymphangiectasias with a JI of 81 ± 6 %. The other lesion types were: angioectasias (JI 64 ± 11 %), aphthae (JI 64 ± 8 %), chylous cysts (JI 70 ± 14 %), polypoid lesions (JI 75 ± 21 %), and ulcers (JI 56 ± 9 %). CONCLUSION: MLA can perform as well as human readers in the measurement of SB angioectasias in white light (WL). Automated lesion measurement is therefore feasible. KID is currently the only open-source CE database developed specifically to aid development of MDSS. Our experiment demonstrates this potential.