Unsupervised Deep Learning Framework for Accelerated LGE MRI Acquisition with Enhanced Image Reconstruction

Motivation: 3D Late Gadolinium Enhancement (LGE) MRI of the left atrium often has variable quality and can be a very long acquisition. Goal(s): Lowering the acquisition time of LGE MRI scan, without sacrificing the final reconstruction quality, and while obtaining a high resolution isotropic 1.25mm dataset. Approach: We used a variable density acquisition and reconstructed the first ~3 minutes of data of the study with a Self-guided Deep Image Prior Network. Reconstructions were compared to a Compressed Sensing iterative reconstruction of a four times longer scan. Results: The CNN-based Deep Image Prior Network produced clinically diagnostic quality reconstruction with 1/4th of the acquisition time. Impact: The network enables high-quality cardiac LGE imaging in 2-4 minutes, representing a significant advancement in clinical workflow efficiency and patient comfort. The reduced acquisition time could expand the accessibility of 3D LGE studies of atrial fibrillation patients.