Image Quality Transfer Using Generative Models

点击翻译按钮获取中文摘要

1. Brief description of the context of the research including potential impact High-field MRI scanners are crucial for accurate diagnostic and clinical management in medical imaging. However, low-field MRI scanners, which have a lower magnetic strength (<1T) compared to 1.5T or 3T scanners, are still widely used in applications such as portable MRI or in lower- and middle-income countries. These scanners often have a lower signal-to-noise ratio (SNR) and less contrast between brain tissues, limiting their application in various clinical stages and for further image analysis. To address these limitations, researchers have proposed using Image Quality Transfer (IQT) to improve the quality of low-field medical images by transferring the rich information from high-field images. Previous approaches to IQT have utilized deep learning to restore high-quality information and have been shown to be superior to other methods, including interpolation and classical machine learning. However, these approaches were only limited to restoring up to x4 image resolution and none of the previous work has tested their methods on clinically relevant metrics or various domain-shifted unseen data, such as lesions and MRI scans from children. 2. Aims and objectives The aims and objectives of this project are:-To explore and use the latest advance in generative models for IQT - Propose more clinically relevant image quality assessment method - Increase computational speed and robustness of the model to unseen/domain-shifted data 3. Novelty of the research methodology In this project, we will propose a new approach for IQT based on recent paradigms in image generation/translation, such as Diffusion models and GANs. This methodology will be validated on various domain-shifted data, such as lesions and MRI scans from children, using our new image quality assessment metric that is more clinically relevant. Furthermore, the PhD will also investigate the application of data-efficient techniques, such as unsupervised/self-supervised learning, to IQT as well as cross-modality. 4. Alignment to EPSRC's strategies and research areas This project strongly fits to the following EPSRC's research themes: - Healthcare Technologies - Information and communication technologies - Artificial Intelligence Technologies 5. Any companies or collaborators involved None