Deep learning with simulated data for medical imaging

This thesis addresses critical challenges in the application of deep learning to medical imaging, particularly focusing on brain vasculature. Deep learning, a specialized branch of machine learning, has shown remarkable potential in tasks such as object classification, detection, and image segmentation. However, its effectiveness is often limited by the scarcity of large, diverse, and labeled datasets in the medical domain. This scarcity stems from the time-consuming and expertise-dependent nature of medical image annotation, making it impractical to manually create extensive datasets for training deep learning models. The research tackles three primary challenges: the disparity in data availability across different imaging modalities, the
underrepresentation of certain phenotypes in medical imaging datasets, and the limited availability of data for rare conditions. To address these issues, the thesis explores innovative approaches to data generation and synthesis, aiming to augment existing datasets and create new ones.
The work employs advanced techniques such as cross-modality synthesis using learned local attention masks to generate MRA images from T2-weighted brain MR images, addressing the data availability discrepancy between modalities. Furthermore, the thesis investigates the use of diffusion models for synthetic generation of brain vasculature, particularly focusing on the intricate Circle of Willis to address phenotype underrepresentation.
The research also introduces few-shot learning with diffusion models to enable conditional generation of brain vessels with aneurysms, tackling the
challenge of limited training data for rare conditions.
By systematically addressing these challenges, this thesis contributes to advancing the field of medical imaging and enhancing the training of deep
learning models in healthcare applications.