SeamXSim: Seamless-textured virtual colonoscopy simulator via unpaired long-term video translation.

Abstract

Colonoscopy is a critical procedure for diagnosing and treating gastrointestinal diseases, but its complexity requires extensive training for proficiency. Complication rates are frequently higher and procedure times frequently longer when colonoscopy is performed by novice endoscopists, indicating a need for effective training tools like virtual simulators. The unrealistic colors and textures of current simulators, however, diminish immersion, thereby limiting their practical utility. To address this, we propose SeamXSim, a seamless-textured colonoscopy simulator that synthesizes real colon images to generate high-fidelity, diverse textures, bridging the gap between virtual and real procedures. Additionally, SeamXSim-T extends this framework with a video translation method that incorporates a novel flow-warp loss, enabling the generation of long-term, temporally consistent video sequences of the entire colon for comprehensive simulation. Experimental results showed that SeamXSim outperformed existing models, achieving a lower Mesh-to-Cloud error (3.629 mm vs. 5.584 mm) in 3D reconstruction tasks and an error rate of 0.123 in depth estimation. SeamXSim-T further strengthened spatial consistency in video translation tasks by 16%. A preference test with novice endoscopists confirmed its improved realism and educational value. In conclusion, SeamXSim can significantly enhance the realism and fidelity of virtual colonoscopy simulations, offering strong potential for enhancing clinical education and improving the skills of novice endoscopists.