Background: Narrow band imaging (NBI) can assist endoscopists in detecting early gastric cancer (EGC) more easily, but its widespread use is hindered by economic cost and technical property rights. We aim to realize the conversion of white light endoscopy (WLE) images into virtual narrow band imaging (Vir-NBI) images using stable diffusion.
Methods: Endoscopic images were retrospectively collected from 325 patients who underwent endoscopic submucosal dissection (ESD). A total of 273 NBI images from 218 patients were used to fine-tune stable diffusion, which then converted 111 WLE images from 107 patients into Vir-NBI images. Endoscopists assessed the images and evaluated their effectiveness in diagnosing EGC and depicting lesion margins in the form of WLE, NBI, and Vir-NBI image pairs.
Results: Compared with WLE images, Vir-NBI images have better quality. The accuracy of junior endoscopists in diagnosing EGC by observing WLE images alone, simultaneous WLE and NBI images, and simultaneous WLE and Vir-NBI images were 61.26%, 79.28% and 81.08%, respectively. For intermediate endoscopists, the diagnostic accuracy was 72.07%, 86.79% and 84.68%, respectively. For senior endoscopists, the diagnostic accuracy was 80.18%, 95.50% and 92.79%, respectively. In addition,Vir-NBI images had higher area concordance rate andsuccessful whole-lesion diagnosis than WLE images (43.85% vs 39.32%, p < 0.001) (45.33% vs 32.87%, p < 0.001).
Conclusions: Vir-NBI images has similar observation effect with real NBI image, which helps endoscopists better visualize the lesion structure, thus improving the accuracy of EGC diagnosis.
Keywords: Artificial intelligence; deep learning; early gastric cancer; narrow band imaging; stable diffusion.