The authors of the Kvasir-Instrument dataset highlight the need for improved tracking and analysis of gastrointestinal (GI) procedures, particularly during and after colonoscopies. The dataset comprises 590 annotated frames featuring GI procedure tools like snares, balloons, and biopsy forceps. It includes ground truth masks and bounding boxes, validated by two expert GI endoscopists. The authors emphasize that critical information, such as disease borders, development, and the size of resected areas, often goes untracked, leading to challenges in follow-up and post-treatment reassessment. Additionally, they provide a baseline for GI tool segmentation to promote algorithm development.

In the context of minimally invasive surgery (MIS), the authors acknowledge its advantages, such as reduced pain, shorter hospital stays, faster recovery, and less scarring compared to traditional open surgery. They note the complexity of MIS procedures, where surgeons must navigate precise hand-eye coordination, often resulting in limited mobility and a restricted field of view. Importantly, they highlight that, unlike accessory organ treatments, GI tract procedures don’t require incisions, making them less invasive. GI procedures encompass both surveillance and treatment, involving various tools like balloon dilatation, biopsy forceps, snares, and submucosal injections. The authors suggest that computer and robotic-assisted surgical systems can enhance surgical capabilities by providing additional patient information for decision-making. However, they stress the need to understand the spatial relationships between surgical instruments, cameras, and anatomy, particularly in GI tract endoscopy.

Kvasir-Instrument labels were made using HyperKvasir unlabelled data. The dataset was sourced from endoscopic examinations conducted at Bærum Hospital in Norway, using images and videos captured with Olympus and Pentax endoscopy equipment. They emphasize that all data collection adhered to patient consent protocols, and patient information was not used for archiving. Anonymization measures, such as random naming for publicly released frames, were implemented.

Regarding the annotation strategy, the authors describe uploading the dataset to labelbox and labeling the Region of Interest (ROI) for diagnostic and therapeutic tools, generating ground truth masks. They provide examples of annotated images, bounding boxes, image annotations, and generated masks:

The First two rows represent frames within biopsy forceps, the middle row consists of the metallic clip, the fourth row is a radiofrequency ablation probe and the last row depicts the crescent and hexagonal-shaped snares for polyp removal.

Annotation was performed in a three-step strategy: initial labeling by two experienced research assistants, cross-validation by two experienced GI experts, and final validation based on expert feedback. The Kvasir-Instrument dataset encompasses 590 frames featuring various GI endoscopy tools, annotated with bounding boxes and segmentation masks. Most tools are small to medium-sized, and the dataset is made publicly available, including original images, masks, bounding box information, and a visualization script for researchers to access and use.