Variational Autoencoders (VAEs) have become a widely used deep learning framework for representation learning in biomedical imaging, with most applications relying on two-dimensional (2D) convolutional architectures tailored to standard image data. While effective for many tasks, 2D VAEs are limited in their ability to capture spatial dependencies across volumetric biomedical data, such as MRI or CT scans, which are inherently three-dimensional (3D). To date, 3D VAE implementations and systematic evaluations remain underdeveloped, leaving a gap in leveraging full spatial information in clinical imaging applications.

This thesis aims to address this gap by developing and integrating a 3D convolutional VAE implementation into the group’s autoencoder framework AUTOENCODIX (https://github.com/jan-forest/autoencodix)

Building on the framework’s modular design, the project will extend current functionality to support efficient training and evaluation of 3D architectures on large-scale biomedical image datasets.

The new implementation will be used to investigate two central research questions:

1. Computational performance optimization: How can 3D VAEs be trained efficiently on large-scale biomedical image data, considering memory constraints and computational overhead?
2. Embedding quality and clinical relevance: How does the embedding performance of 3D VAEs compare to established 2D VAEs, particularly when applied to clinical imaging data from stroke patients?

Through these investigations, the thesis will provide insights into both the technical feasibility and biomedical utility of 3D VAE embeddings. The work will contribute to advancing deep generative modeling in clinical imaging by bridging the methodological gap between 2D-focused approaches and the growing need for 3D representation learning in translational medicine.

Students profile: Master Data Science/Bioinformatik/Informatik with knowledge and experience in Python Programming and PyTorch implementation. Basic understanding of medical imaging and analysis.