Abstract

Gastrointestinal (GI) pathologies are periodically screened, biopsied, andresected using surgical tools. Usually the procedures and the treated orresected areas are not specifically tracked or analysed during or aftercolonoscopies. Information regarding disease borders, development and amountand size of the resected area get lost. This can lead to poor follow-up andbothersome reassessment difficulties post-treatment. To improve the currentstandard and also to foster more research on the topic we have released the``Kvasir-Instrument'' dataset which consists of $590$ annotated framescontaining GI procedure tools such as snares, balloons and biopsy forceps, etc.Beside of the images, the dataset includes ground truth masks and boundingboxes and has been verified by two expert GI endoscopists. Additionally, weprovide a baseline for the segmentation of the GI tools to promote research andalgorithm development. We obtained a dice coefficient score of 0.9158 and aJaccard index of 0.8578 using a classical U-Net architecture. A similar dicecoefficient score was observed for DoubleUNet. The qualitative results showedthat the model did not work for the images with specularity and the frames withmultiple instruments, while the best result for both methods was observed onall other types of images. Both, qualitative and quantitative results show thatthe model performs reasonably good, but there is a large potential for furtherimprovements. Benchmarking using the dataset provides an opportunity forresearchers to contribute to the field of automatic endoscopic diagnostic andtherapeutic tool segmentation for GI endoscopy.