MultiPRESS: Deep learning for super-resolution 4D Flow MRI
Title: |
MultiPRESS: Deep learning for super-resolution 4D Flow MRI |
DNr: |
Berzelius-2022-253 |
Project Type: |
LiU Berzelius |
Principal Investigator: |
David Marlevi <david.marlevi@ki.se> |
Affiliation: |
Karolinska Institutet |
Duration: |
2023-01-16 – 2023-03-31 |
Classification: |
20604 |
Homepage: |
https://staff.ki.se/people/david-marlevi |
Keywords: |
|
Abstract
Medical imaging has come to revolutionize the way we diagnose, monitor, and treat disease. This is particularly true in the field of cardiovascular care, where a range of imaging modalities are used to interrogate anatomy and function using non-invasive means. However, a range of clinical challenges remain where available image techniques are hindered by inherent method limitations (restricted resolution; excessive noise; etc.). The incorporation of machine learning techniques and data-driven algorithms into clinical image research is however starting to fundamentally change what can be analysed, and resolved, opening for individualized patient-care and directly improved clinical risk stratification.
The purpose of this proposal is to implement and explore super-resolution neural networks to enhance the spatiotemporal resolution of 4D Flow magnetic resonance imaging (MRI) – a novel imaging technique enabling imaging of 3D blood flow across the entire cardiovascular system. Work will include training of simulated image sets, allowing for controlled implementation and testing of various super-resolution networks across a range of architectures (ResNet-CNN; GAN; DRL; PINN). Work may expand to include either exploration of alternative or adjusted architectures, or incorporation of physics-informed loss function variations. The work may also include processing of anonymized clinical image data, with validation and testing expected in a direct clinical setting on patients with disease types relevant for super-resolution flow imaging.