Multifrequency Time-Dependent Deep Image Prior for Real-Time Free-Breathing Cardiac Imaging

Abstract

The aim of this study is to enable high temporal resolution functional cardiac imaging without breathholds or electrocardiogram (ECG) gating. Real-time MRI is essential for assessing heart function in patients with limited breathhold capacity or arrhythmias that preclude breathheld ECG-gated cine scans. The Time-Dependent Deep Image Prior (Time-DIP) method is a promising reconstruction for dynamic MRI, combining a nonlinear manifold with zero-shot deep learning. However, while Time-DIP has been demonstrated for breathheld cine imaging, it employs a helical manifold that assumes quasi-periodic motion and thus may not be suitable for free-breathing real-time scans, particularly in arrhythmia patients. This study proposes a Multifrequency Time-DIP technique to extend this framework to free-breathing real-time cardiac imaging. First, a “multifrequency manifold” is introduced that parameterizes time using multiple sinusoids spanning various frequencies, enabling dynamic imaging without assuming motion periodicity. Second, joint estimation of coil sensitivities using zero-shot deep learning is used to improve the reconstruction of multichannel data. Simulations and scans of healthy subjects and patients, including those with arrhythmias, were performed using a 2D free-breathing ungated golden angle spiral bSSFP sequence. Image quality and left ventricular (LV) functional measurements were compared to real-time scans reconstructed with compressed sensing and the original Time-DIP implementation, as well as conventional breathheld ECG-gated cine scans. Multifrequency Time-DIP outperformed other real-time techniques in simulations of various motion scenarios. In vivo scans using Multifrequency Time-DIP exhibited reduced aliasing artifacts, achieving temporal resolutions as high as a single TR (4.2 ms/frame), with no significant differences in LV functional measurements compared to conventional scans. While conventional scans had better edge sharpness and image contrast scores, Multifrequency Time-DIP exhibited overall higher image quality metrics among real-time scans. In conclusion, a generalized Time-DIP reconstruction was shown to enable high temporal resolution free-breathing real-time cardiac imaging in healthy subjects and patients, including those with arrhythmias.