Master's Thesis presentation, Simon Carlson

Abstract: In this research, we introduce an innovative approach by seamlessly merging Variational Autoencoders (VAEs) with Nonlinear Mixed-Effects (NLME) models. Our primary objective is to generate a synthetic dataset that ties medical images to their corresponding synthetic longitudinal data. The study unfolds through three pivotal stages. First, we employ a VAE to transform medical images into a latent space representation. Next, we establish a distinct association between the VAE's latent space and the NLME model's random effects. This intricate connection facilitates the creation of a direct link between synthetic images and longitudinal data, positioning the latent variables as essential connectors. Finally, we implement a comprehensive validation process to evaluate our approach's efficacy.

Our results underscore the robustness of the integrated framework, demonstrating a profound interrelation between synthetic images and their associated longitudinal data. This research indicates that medical images can serve as potent predictors for synthetic longitudinal outcomes. Furthermore, the study emphasizes that integrating images remains advantageous, even in the presence of noise.

Furthermore, by generating synthetic datasets, we offer a solution that addresses critical issues of data privacy and financial constraints, presenting a practical and enticing option for researchers to develop and test their models without compromising patient information. Ultimately, this study posits that the integration of VAEs with NLME models, applied to medical imaging data, could pave the way for more accurate, timely, and personalized medical care by utilizing images beyond their traditional diagnostic role.