Generation of Synthetic Data for Medical Decision Support Applications

Computer vision has the potential to accelerate decision support in a variety of medical applications, but a paucity of high-quality, open-source datasets hinders the development of decision support applications for open medical environments (i.e., non-laparoscopic). Synthetic data holds promise as a solution for difficult-to-obtain data, such as images from high stress environments or complicated by privacy concerns associated with medical imagery. Modern applications of synthetic data generation such as digital twins (DT) can create high-fidelity and photo-realistic images the surpass traditional data augmentation practices. Moreover, synthetic data is cost-effective, efficient, and highly scalable after initial creation of the assets and environment. This study presents a framework to synthetically generate annotated images from digital 3-D assets in random perspectives and orientations for use in computer-vision (CV) for open medical applications. Datasets of 1,000 annotated synthetic images and 681 real-world images of six object classes were employed in a transfer learning toolkit for an experiment to determine the feasibility of utilizing synthetic data to train models on CV applications to augment or replace training with real-world data. The experiment tested model performance based on three datasets: 1. Synthetic data only; 2. Real-world data only; 3. Training with synthetic data for evaluation with real-world data. Results showed that training on synthetic data to evaluate real-world data met or exceeded the performance of training on real-world data in four of six classes within the datasets utilized. Results show promise for utilizing synthetic data as an alternative to costly, time consuming, and difficult to obtain data types with many areas for further study, such as more detailed and comprehensive environments and assets as well as methodology for noise injection to improve model performance.