Construction and validation of a high-precision annotated dataset for developing intelligent critical vein recognition models in laparoscopic pancreatic surgery

TL;DR

This paper creates and validates a dataset for AI models to identify critical veins during laparoscopic pancreatic surgery, aiming to improve surgical precision and training.

Contribution

The first large-scale, expert-annotated dataset for vein recognition in laparoscopic pancreatic surgery is constructed and made publicly available.

Findings

The dataset contains 19,003 annotated frames with precise vein segmentations.

A baseline model achieved 79.6% recall, 95.8% precision, and a Dice coefficient of 0.69 on the test set.

The dataset and benchmark provide a foundation for advancing AI-assisted vascular identification in laparoscopic surgery.

Abstract

Laparoscopic operation holds multiple advantages as a minimal invasive method of surgical treatment. Vascular-related manipulations, including identification and dissection of vascular structures and control of bleeding, are highly experience-based and demand a tortuous learning curve. With the rapid development of artificial intelligence (AI) in the entire diagnosis and treatment process of diseases, data-driven AI models have shown promising potentials in both education and real-time aiding in surgery. However, there is no dedicated dataset existing for developing vascular identification models in laparoscopic settings.

Videos from 23 laparoscopic distal pancreatectomy (LDP) and laparoscopic pancreaticoduodenectomy (LPD) performed at Peking Union Medical College Hospital (PUMCH) between January 2021 and June 2022 were collected. Senior surgeons systematically reviewed surgical videos to visually identify critical venous vasculature, precisely annotating frame-accurate start and end timestamps on the video timeline. Frames were extracted from these video segments at a fixed temporal interval of one frame per second, then stored to compile the source image dataset. The contours of superior mesenteric vein (SMV), portal vein (PV), splenic vein (SV) were labelled and reviewed according to standard procedure. A High-Resolution Network (HRNet) was combined with a fully convolutional network (FCN) output head to construct a preliminary segmentation model for initial validation and comparison.

A dataset comprises 19,003 annotated frames and is publicly available. The baseline model achieved a recall of 79.6%, precision of 95.8%, and Dice coefficient of 0.69 on the testing set.

This study constructed and released the first large-scale, expert-annotated dataset of key venous structures from pancreatic surgery (PS) videos and established benchmark performance for intraoperative vein segmentation using open-source models. This resource provides a foundation for advancing AI-assisted vascular identification in laparoscopic surgery.

Figures