EIVS: Unpaired Endoscopy Image Virtual Staining via State Space Generative Model

In recent years, gastrointestinal diseases have posed a growing threat to global human health. Endoscopy is a crucial method for diagnosing gastrointestinal diseases, but small lesions can be difficult to be observed with standard white light endoscopy (WLE). Chromoendoscopy, which involves staining tissues with dye to highlight differences between diseased and normal tissues, has been widely adopted to improve diagnosis. However, the effectiveness of chromoendoscopy is highly dependent on the expertise of medical practitioners, and the dyes can potentially cause adverse reactions in patients. To address these limitations of chromoendoscopy, we proposed a new model called EIVS, for virtual staining of WLE images. We designed Skip State Space Connection (TSC) in EIVS, which is skip connection incorporating State Space Models. TSC captures long-range contextual relationships and facilitates multiscale feature fusion. Experimental results on a clinical unpaired endoscopy-chromoendoscopy dataset indicate that our proposed model generates high quality virtual chromoendoscopy images, and outperforms other state-of-the-art methods.