Controllable synthetic algorithms and evaluations for annotated clinical image synthesis

Abstract

Medical imaging is an essential tool in healthcare, offering vital insights for diagnosis and treatment. These non-invasive techniques, from X-rays to advanced MRI scans, are crucial in clinical practice but face challenges like data availability, privacy concerns, and the need for accuracy. This thesis explores the potential of synthetic data in medical imaging as a promising solution to these challenges. Using advanced deep learning methods, synthetic medical images can supplement real datasets, ensuring privacy and enhancing the diversity of medical conditions for study. The research focuses on developing and applying algorithms for medical image synthesis to improve data analysis quality and efficacy.

Chapter 2 focuses on synthesizing image data for complex modalities with restricted protocols. The study enhances the temporal resolution of 3Dir MVM CMR, a cardiac MRI technique, using a hybrid UNet and Generative Adversarial Network.

Chapter 3 introduces an Unsupervised Mask (UM)-guided synthesis strategy for generating medical image annotations. This method outperforms previous techniques in creating high-fidelity, diverse, and useful synthetic CT images.

Chapter 4 presents new evaluation methods for synthetic images, examining fidelity, diversity, privacy, and utility. The study finds a balance between these factors and shows that images with lower fidelity can still be useful for tasks like data augmentation.

Chapter 5 discusses the practical use of synthetic medical images, highlighting their role in patient privacy and enhancing data for medical tasks. Evidence shows synthetic images can forecast mortality in specific patient groups, demonstrating their clinical value.

In summary, this thesis presents a comprehensive exploration of the use of synthetic data in medical imaging. Through advanced deep learning techniques, it demonstrates how synthetic medical images can augment existing datasets, addressing key challenges such as data scarcity, privacy concerns, and the need for high-accuracy diagnostic tools.