CEID: Benchmark Dataset for Designing Segmentation Algorithms of Instruments Used in Colorectal Endoscopy

Recently, a large number of computer-assisted methods have been reported for colorectal lesions (e.g., polyp, inflammation, perforation, etc.) detection, segmentation, and classification in endoscopy to improve the operating efficiency of clinicians and surgical robots. However, few works specially involve endoscopic instruments which play a great role in lesion screening, tracking, and diagnosis. To promote the development of this field, we create a colorectal endoscopic instrument dataset (CEID) in this paper, which consists of 1032 images containing colorectal procedure instruments, such as snares, titanium clips, entry needles, high frequency electrotomes and biopsy forceps. The segmentation mask of each image was labeled and verified by two experienced gastroenterologists. Furthermore, we selected 5 classic general-purpose segmentation algorithms built for the medical image segmentation task and 2 specific-purpose segmentation algorithms built for the digestive lesion segmentation task, and tested their performance on the collected dataset. Experimental results demonstrate that, our dataset is capable of verifying different segmentation networks, and both kinds of methods cannot fully meet the clinical needs, leaving a large potential for further improvements. Benchmarking using the dataset provides an opportunity for researchers to contribute to the field of automatic instrument segmentation in colorectal endoscopy.