Rosette Cardiac MR Fingerprinting for Simultaneous T1, T2, T 2 * , and Fat Fraction Mapping Using a Multi-Echo Deep Image Prior Reconstruction

Abstract

Purpose: Quantitative mapping of cardiac tissue properties is used clinically in diagnosis and monitoring of a wide variety of cardiac pathologies. Cardiac Magnetic Resonance Fingerprinting (cMRF) enables rapid and simultaneous quantification of multiple parameters in the myocardium from a single scan. In this work, a multi-echo cMRF acquisition is combined with a deep image prior framework to reconstruct cardiac T₁, T₂, $T_2^{*}$ , and fat fraction maps.

Methods: A 2D, single-breathhold, ECG-gated rosette trajectory cMRF sequence was deployed to sensitize the signal to T₁, T₂, $T_2^{*}$ , and fat off-resonance effects. Data were processed using a deep image prior reconstruction trained with the cMRF encoding model to generate images consistent with the acquired k-space data. These images were used in curve fitting and pattern matching algorithms to generate T₁, T₂, $T_2^{*}$ and fat fraction maps. The technique was validated using numerical simulations, standard phantoms, and 28 healthy subjects.

Results: In phantoms, good agreement was observed between the proposed technique and gold-standard reference measurements. In healthy subjects, measurements made with the deep image prior (DIP) reconstruction agreed with clinical cardiac measurements and demonstrated smaller voxel-level variance in a healthy population compared to iterative low-rank and direct matching reconstructions.

Conclusion: The multi-echo cMRF acquisition coupled with a DIP reconstruction enables the simultaneous quantification of T₁, T₂, $T_2^{*}$ , and fat in the heart and demonstrates good agreement with conventional mapping approaches in phantom and in vivo experiments. Additionally, the DIP reconstruction provides accurate measurements with a lower voxel-level variance compared with direct gridding and iterative low-rank reconstruction methods.

Keywords: MR fingerprinting; cardiac MRI; deep learning reconstruction; quantitative MR.