SimIntestine: A synthetic dataset from virtual capsule endoscope

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Accurate position-annotated datasets are crucial for deep learning models in gastrointestinal (GI) endoscopy.
• Existing synthetic datasets lack realistic anatomical and textural features of the GI tract.
• Challenges exist in obtaining annotated data from real endoscopes for pose and depth estimation.

Purpose of the Study:

• To develop a method for generating a realistic, position-annotated synthetic dataset of the human intestines.
• To provide ground truth depth maps and camera pose information for training and evaluating deep learning models.
• To address the limitations of current datasets for endoscopic video analysis.

Main Methods:

• Generation of virtual models of the small and large intestines with realistic anatomical features (e.g., plicae circulares, villi, haustral folds) and textures.
• Integration of a virtual capsule endoscope simulating real-world navigation and image capture.
• Incorporation of physiological processes like peristalsis into the virtual environment.
• Generation of position-annotated images, ground truth depth maps, and camera orientation/position data.

Main Results:

• The SimIntestine dataset provides physically realistic, annotated synthetic data of the human GI tract.
• The dataset includes distinctive anatomical features and simulates physiological processes.
• Evaluated benchmark techniques (Endo-SfMLearner, Monodepth2) using SimIntestine, demonstrating its efficacy through enhanced performance metrics.

Conclusions:

• SimIntestine serves as a valuable benchmark for improving endoscopic video analysis, particularly for pose estimation and simultaneous localization and mapping.
• The generated dataset overcomes the limitations of real-world unannotated data.
• Quantitative evaluation confirms the dataset's efficacy and its ability to enhance model performance.