Generalized Super-Resolution 4D Flow MRI - Using Ensemble Learning to Extend Across the Cardiovascular System.

Area of Science:

• Medical Imaging
• Cardiovascular Science
• Artificial Intelligence

Background:

• 4D Flow MRI quantifies cardiovascular blood flow non-invasively.
• Limitations include spatial resolution and image noise.
• Super-resolution (SR) networks can improve image quality post-scan.

Purpose of the Study:

• To explore the generalizability of SR 4D Flow MRI across diverse cardiovascular domains.
• To evaluate ensemble learning techniques for SR domain adaptation in 4D Flow MRI.
• To assess SR performance using heterogeneous training sets and various network architectures.

Main Methods:

• Generated synthetic data across cardiac, aortic, and cerebrovascular domains.
• Trained and evaluated existing SR base models and ensemble learners (bagging, stacking).
• Quantified performance using in-silico and in-vivo data from the three domains.

Main Results:

• Ensemble methods (bagging, stacking) significantly enhanced SR performance across domains.
• Accurate prediction of high-resolution velocities from low-resolution data in-silico.
• Successful recovery of native velocities from downsampled in-vivo data and potential for denoising.

Conclusions:

• Ensemble learning offers a viable approach for generalized SR 4D Flow MRI.
• This method extends the utility of SR across various clinical cardiovascular applications.
• Novel application of ensemble learning to advanced full-field flow imaging.