With advances in digital healthcare technologies, optional therapeutic modules and tasks such as depth estimation, visual localization, active control, automatic navigation, and targeted drug delivery are desirable for the next generation of capsule endoscopy devices to diagnose and treat gastrointestinal diseases. Although deep learning applications promise many advanced functions for capsule endoscopes, some limitations and challenges are encountered during the implementation of data-driven algorithms, with the difficulty of obtaining real endoscopy images and the limited availability of annotated data being the most common problems. In addition, some artefacts in endoscopy images due to lighting conditions, reflections as well as camera view can significantly affect the performance of artificial intelligence methods, making it difficult to develop a robust model. Realistic simulations that generate synthetic data have emerged as a solution to develop data-driven algorithms by addressing these problems. In this study, synthetic data for different organs of the GI tract are generated using a simulation environment to investigate the utility and generalizability of the synthetic data for various medical image analysis tasks using the state-of-the-art Endo-SfMLearner model, and the performance of the models is evaluated with both real and synthetic images. The extensive qualitative and quantitative results demonstrate that the use of synthetic data in training improves the performance of pose and depth estimation and that the model can be accurately generalized to real medical data.
Key words: Synthetic data generation, capsule endoscopy, depth and pose estimation
|
