A tool for generating and evaluating synthetic embryo development videos using generative models

Abstract There has been a notable increase in the use of generative models over the past few years. However, there is still a lack of full exploitation due to the difficulty of their implementation, particularly in the medical field. This work presents a tool for generating synthetic videos of embryo development and an analysis to identify the most suitable generative model for this task. The proposed tool streamlines the use of generative models for technical personnel by enabling the generation of multiple video versions for the same case, automatically evaluating their quality, and assigning scores. In addition to generating synthetic videos, this tool provides automatic analysis of the output, ranking the videos based on their resemblance to real cases and absence of inconsistencies or noise. This automatic analysis ensures a higher quality and more realistic portrayal of embryo development, while making advanced generative models more accessible to a wider audience. Moreover, a comprehensive examination of two distinct generative models, one based on Generative Adversarial Networks and another based on Diffusion Models, reveals that the Diffusion Model excels at precisely locating embryo stages in more accurate positions, resulting in more realistic videos.