AI-driven cardiovascular imaging
Abstract
Learning Data-Driven Representations of Cardiovascular Velocity Fields from 4D Flow MRI
This project aims to discover data-driven features of the cardiovascular velocity field from 4D Flow magnetic resonance imaging using representation learning. 4D Flow MRI consists of time-resolved, three-dimensional, three-directional magnetic resonance flow imaging and generates large, high-dimensional data with millions of data points. This technique is applied throughout the world to assess cardiovascular blood flow in vivo in health and disease. Current analyses rely largely on predefined hemodynamic biomarkers, such as peak velocity, helicity, wall shear stress, or turbulent kinetic energy. While informative, these handcrafted metrics capture only selected aspects of the highly complex spatiotemporal blood flow field. Representation learning offers the possibility of extracting latent features directly from the data, potentially revealing previously unrecognized hemodynamic phenotypes associated with disease mechanisms, progression, and clinical outcomes.
We developed a prototype 4-channel dual-encoder autoencoder that reconstructs anatomical and velocity images of 486 4D Flow datasets of the carotid arteries. Training three-dimensional autoencoder architectures on time-resolved 4D Flow MRI data is computationally demanding due to the size of the datasets and the memory requirements associated with volumetric spatiotemporal convolutions. Access to Berzelius enables efficient large-scale hyperparameter optimization, training with larger batch sizes, and development of more advanced architectures that would not be feasible on local workstations
