Fusing Deep Learning Based Resolution Enhancement and Ultrahigh Field MRI for High Resolution Anatomical and Structural Imaging of the Human Heart

• Applicant: Dr. David Lohr
• Subject Area: Medical Physics, Biomedical Technology ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Radiology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 519067094

Low signal-to-noise ratio (SNR) at 1.5T and 3T limits the application of cardiac diffusion tensor imaging (cDTI) and high-resolution CINE imaging in clinically reasonable scan times. Ultrahigh field (UHF) MRI has been shown to offer increased SNR and achievable image resolutions in both neurological and cardiac MRI. Primary aim of this study is to harness the SNR gain associated with UHF (7T) MRI in combination with deep learning approaches to enable cDTI and high-resolution CINE imaging in clinically reasonable scan times. Intermittent objectives towards this aim are 1) the training, validation, and testing of resolution enhancement models based on publicly available neurological 7T data 2) the implementation and validation of a cDTI sequence at both 3T and 7T 3) the training, validation, and testing of resolution enhancement models based on cardiac locally acquired 7T CINE and cDTI data 4) the assessment of generalization capabilities of the developed models with respect to different field strengths (7T to 3T) and organs (brain to heart) as well as disease (hypertrophic cardiomyopathy and amyloidosis). While cardiac UHF MRI is currently rather a research modality, neurological UHF MRI is well established and clear clinical benefits have been described. We therefore aim to establish resolution enhancement models using 7T brain images from the publicly accessible human connectome project repository in a first step. Respective models will be based on UNet architectures in combination with 1) generative adversarial networks and 2) perception or feature loss. I hypothesize that both cardiac applications are ideal for deep learning based resolution enhancement models, because they rely on the acquisition of a large number of images. In the second step, established models and methodology will thus be transferred and adopted towards 7T cDTI and CINE data. The combination of high-resolution imaging at 7T with deep learning based resolution enhancement may fundamentally change how future clinical images will be generated. The concept of resolution enhancement would enable the transfer of information from expensive, high quality images, to low cost, low quality images, for example from 7T to 3T or even 1.5T for potentially all organs. Clinical benefits associated with 7T would become accessible to a larger patient population, increasing the diagnostic value of MRI, while simultaneously improving patient comfort via shorter examination times.

• DFG Programme: Research Grants