HiEndo: harnessing large-scale data for generating high-resolution laparoscopy videos under a two-stage framework

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Surgical Simulation

Background:

• Generative AI shows promise in medicine, but realistic gastrointestinal laparoscopy video generation remains underexplored.
• Existing models like Endora produce low-resolution videos insufficient for clinical needs in robotic surgery.

Purpose of the Study:

• To propose HiEndo, an innovative two-stage architecture for generating high-resolution, high-fidelity gastrointestinal laparoscopy videos.
• To overcome the limitations of prior methods in generating clinically relevant surgical videos.

Main Methods:

• A two-stage approach: first, a diffusion transformer builds upon Endora for initial low-resolution video generation.
• Second, a super-resolution module enhances video resolution and refines fine-grained details.
• A large-scale dataset of 61,270 gastrointestinal laparoscopy video clips was curated for training and validation.

Main Results:

• The HiEndo framework successfully generates high-resolution, realistic gastrointestinal laparoscopy videos.
• Achieved significant improvements over state-of-the-art methods, including a 15.1% reduction in Fréchet Video Distance and a 3.7% F1 score increase.
• Experimental results validate the effectiveness of the proposed two-stage generation architecture.

Conclusions:

• HiEndo provides a viable solution for generating high-fidelity gastrointestinal laparoscopy videos.
• The generated videos meet the requirements for real-world clinical usage and surgical training.
• This work advances AI-driven medical video generation for robotic surgery applications.