Generating Synthetic CT Data Using a Diffusion-Based Generative AI Model

Abstract

This master’s thesis addresses the generation of synthetic computed tomography (CT) data using diffusion-based generative models. The motivation for this research is rooted in the challenges frequently encountered in medical image analysis, which often faces issues of data scarcity, stringent privacy regulations, and the significant costs associated with expert annotation. Synthetic datasets represent a promising alternative to mitigate these challenges by providing realistic, privacy-preserving imaging data for research, development, and training purposes. In collaboration with CADS GmbH, Perg, a diffusion framework based on a 3D U-Net architecture was implemented and evaluated. The system was designed to generate volumetric head CT scans conditioned on anatomical label maps, with outputs exported in standard medical imaging formats to ensure compatibility with downstream pipelines. The experimental results demonstrate the feasibility of diffusion models in this context by showing that the proposed model can synthesize anatomically plausible and diverse CT volumes. However, the study also highlights several critical limitations. Most notably, insufficient image contrast and restricted spatial resolution currently prevent direct applicability of the generated data for tasks such as training data enrichment or data augmentation. Furthermore, the high computational cost of volumetric diffusion training, together with the limited diversity of available datasets, remains a major challenge. Nevertheless, this work demonstrates a viable proof of concept for diffusion-based CT synthesis, highlighting both the potential benefits and the existing limitations of this approach. The study identifies several promising directions for further research, such as adopting latent diffusion methods to reduce resource requirements, applying postprocessing techniques for resolution enhancement, and expanding training datasets to improve generalizability and applicability of synthetic CT data.

Date of Award	2025
Original language	English
Supervisor	Herwig Mayr (Supervisor)