A Benchmark Dataset of Endoscopic Images and a Novel Deep Learning Method to Segment Gastric Intestinal Metaplasia Under Linked Color Imaging

ABSTRACT Endoscopy has become a cornerstone in diagnosing gastrointestinal diseases, notably for gastric intestinal metaplasia (GIM)—a condition that significantly elevates the risk of gastric cancer. Variability in medical expertise and patient‐specific complexities render sole reliance on clinicians for GIM diagnosis a process that is both time‐intensive and challenging. There is an urgent clinical need for advanced deep learning methods to assist in clinical decision‐making, particularly in the identification and quantification of GIM areas. Linked Color Imaging (LCI) represents a pioneering advancement in endoscopic technology, offering distinctive color enhancement capabilities. However, the scarcity of GIM datasets acquired using LCI has hindered extensive exploration of tailored deep learning applications. This study directly addresses these dual challenges through two strategic approaches. Firstly, we meticulously assembled a comprehensive and unique dataset—designated as ZD‐LCI‐GIM—comprising 1020 GIM images from 249 patients and 864 non‐GIM images from 303 patients. Each image was acquired via LCI gastroscopy and meticulously annotated by seasoned gastroenterologists—measures that ensure high‐quality data for analysis. Secondly, we introduce VM‐UNetV2—a deep learning model specifically designed for medical image segmentation tasks. This model achieves impressive performance on the ZD‐LCI‐GIM dataset, with a mean Intersection over Union (IoU) of 64.13%, accuracy of 88.79%, F1 score (Dice Similarity Coefficient, DSC) of 78.14%, specificity of 90.96%, and sensitivity of 82.06% for GIM segmentation. Moreover, the model exhibits exceptional efficiency, processing 34.61 frames per second (FPS) when run on an NVIDIA RTX 4090 GPU. In addition, our algorithm demonstrates superior accuracy and competitive efficiency in comprehensive experiments on the public datasets Endoscene, CVC‐ClinicDB, Kvasir, CVC‐ColonDB, and ETIS‐LaribPolypDB. The code for this work is available at https://github.com/nobodyplayer1/VM‐UNetV2 .