Recurrent Generative Adversarial Network (GAN) models for synthetic biological cell dynamics

In the medical field, specifically in biology, extensive datasets are crucial for accurate diagnostics, effective treatments, and robust research. However, the time and cost constraints associated with traditional data collection pose significant challenges. This thesis addresses these challenges by posing initial steps in developing a tool that facilitates data augmentation through synthetic data generation, leveraging Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs). Specifically, Recurrent Generative Adversarial Networks (RGANs) are employed to generate synthetic datasets regarding biological cell dynamics. The reproduced datasets exhibit statistical properties similar to real data extracted from videos of in-vitro experiments provided by the Dipartimento Rizzoli-RIT (Research, Innovation & Technology) of the Istituto Ortopedico Rizzoli. In this work, both tabular features and synthetic videos have been produced, capturing the dynamical and morphological characteristics of biological cells over time. The synthetic videos generated by combining VAEs and GANs resemble those obtained using VAEs without RGANs. As first thesis contribution, a dataset has been extracted from the videos provided by the Dipartimento Rizzoli-RIT of the Istituto Ortopedico Rizzoli, improving its quality by refining the data via segmentation and labeling. The main contribution of this work is the development of a neural network architecture able to produce large volumes of data starting from a limited number of initial videos. The tool developed in this thesis represents a first step toward the development of toolboxes capable of enhancing data augmentation and thus providing additional information for biologists.