Methodology for Generating Medical Images Applied to the Generation Of Synthetic Colon Polyps.

Adenomatous polyps represent the early stages of colorectal cancer cases. They can be detected via colonoscopy procedures, but since it relies on visual analysis, it becomes a challenging task to identify those anomalies, that often vary widely in size and aspect. Deep Learning models represent a great option to contribute with this task, being used in diagnostic aid. However, the lack of available data on this domain restricts further advancements. To address this issue, this paper proposes a new methodology for synthetic data generation of medical images, applied to generate colon images. Four generative models, Guided Diffusion, Poisson Flow Generative Models, Improved Diffusion and Slicing Adversarial Network, where trained to produce colonoscopy exam images containing polyps and their results were analyzed, with the best one achieving Fréchet Inception Distance values of 33.89 and Structural Similarity Index Measure of 0.2573, surpassing the current best state-of-art model for this application. With Precision and Recall for Generative Models of 0.811 and 0.649, respectively, and Kernel Inception Distance of 0.063 it is evidenced that the proposed methodology enables the generation of realistic images of colon polyps, contributing to the increase in the variability of the dataset with synthetic images while preserving the main characteristics of the original distribution.