Minimal data requirement for realistic endoscopic image generation with Stable Diffusion.

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Artificial Intelligence

Background:

• Computer-assisted surgical systems enhance surgical procedure execution and outcomes.
• Deep learning models are crucial for these systems but require extensive, well-annotated data.
• Generating synthetic data is a viable solution to data limitations, but minimizing the domain gap between real and synthetic data is essential.

Purpose of the Study:

• To develop a method for translating synthetic images into realistic ones for training surgical AI.
• To improve the generalizability of deep learning models in computer-assisted intervention guidance systems.

Main Methods:

• A novel image-to-image translation technique based on a Stable Diffusion model is proposed.
• The method utilizes synthetic data as input to generate realistic medical images.
• Control networks are incorporated for finer control over image details and reduced input data requirements.

Main Results:

• The method was successfully applied to laparoscopic cholecystectomy datasets.
• It achieved a mean Intersection over Union (IoU) of 69.76%, significantly outperforming baseline methods (42.21%).

Conclusions:

• The proposed image translation method effectively generates realistic images from synthetic data.
• This advancement facilitates the training of deep learning models that generalize better to real-world surgical scenarios.
• The approach holds promise for improving the performance and reliability of computer-assisted surgical guidance systems.