Generation and evaluation of realistic medical images with latent diffusion models for virtual clinical trials in radiation therapy

Details

Supervisors

Barragan Montero, Ana Maria ; Lee, John Aldo

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée

Abstract

In recent years, deep learning has revolutionized various fields, offering advanced tools for tasks such as generating high-quality synthetic images and performing precise image segmentation. These advancements are particularly impactful in the medical domain, where deep learning algorithms assist in identifying and segmenting tumors, diagnosing diseases,... The effectiveness of these algorithms hinges on the availability of vast amounts of high-quality training data, which is especially critical in medical imaging due to its direct influence on patient outcomes and clinical decisions. However, generating non-synthetic data such as CT scans is resource-intensive, requiring significant financial investment and professional expertise to ensure data quality. Additionally, strict confidentiality regulations, such as the European Union's General Data Protection Regulation (GDPR), raise challenges for researchers in terms of data collection, storage, and processing. These constraints make acquiring adequate datasets for training robust deep-learning models difficult. This thesis addresses these challenges by presenting a comprehensive workflow for generating synthetic lung CT scans with tumors. The proposed workflow integrates two key components: a variational autoencoder to derive latent representations of the CT scans, and a diffusion model utilizing a U-Net architecture trained on these latent representations. The workflow performance is assessed through both quantitative and qualitative evaluations. Furthermore, several modifications are recommended to enhance the workflow potential, aiming to provide a reliable and efficient method for generating high-quality synthetic medical imaging data.