Synthesized colonoscopy dataset from high-fidelity virtual colon with abnormal simulation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning models for colonoscopy require extensive, high-quality image datasets for training.
• Limited availability of real colonoscopy images due to privacy and regulatory issues hinders model generalization.
• Existing datasets often lack diversity and sufficient examples of critical abnormalities.

Purpose of the Study:

• To develop a method for generating high-fidelity 3D colon models and synthesizing diverse colonoscopy images.
• To create a comprehensive dataset of synthetic colonoscopy images with abnormalities (polyps, bleeding, ulcers) for training AI models.
• To evaluate the effectiveness of the synthesized dataset in improving deep learning model performance for colonoscopy tasks.

Main Methods:

• Derived colon geometry from CT images.
• Utilized surface mesh deformation and texture mapping for realistic polyp and ulcer modeling.
• Simulated blood diffusion for realistic bleeding effects and incorporated these into the 3D colon models.

Main Results:

• Generated a comprehensive dataset of high-fidelity, synthesized colonoscopy images featuring various abnormalities.
• Trained state-of-the-art deep learning models on the synthesized dataset.
• Models trained on synthetic data demonstrated enhanced performance in abnormal classification, detection, and segmentation tasks.

Conclusions:

• The proposed method effectively synthesizes realistic colonoscopy images with abnormalities.
• The generated dataset significantly improves the generalization and performance of deep learning models for colonoscopy.
• This approach offers a viable solution to the data scarcity problem in training AI for colonoscopy.