1356032 - Cardiac Phase-Resolved T1, T2, and M0 Mapping and Bright/Dark Blood Cine Imaging using MR Fingerprinting with Self-Supervised Deep Learning

Cine MRF has the potential to shorten exam times by acquiring cine images and quantitative maps during an all-in-one approach (1,2). However, previous work used nonrigid motion correction that may be prone to errors, as the images used for motion estimation often have low SNR, variable contrasts, and aliasing artifacts. This work introduces a self-supervised deep learning reconstruction for 2D cine MRF that yields true cardiac phase-resolved (not motion-corrected) T₁, T₂, and M₀ maps, along with bright and dark blood cine images, at a high temporal resolution in a 5-10s breathhold.


Methods:

This work extends a low-rank deep image prior (DIP) reconstruction described for diastolic MRF (3). A dictionary is simulated using Bloch equations and compressed to K=9 basis signals using the SVD (4). MRF k-space data are binned into P=24 cardiac phases using the ECG. A u-net learns to generate subspace images for each phase as follows. While the original DIP method used a random noise tensor as input (5), here random tensors are initialized for the first and last cardiac phases and those for intermediate phases are linearly interpolated, which imposes some regularization across phases. The u-net is trained by enforcing consistency with the acquired k-space data (Figure 1). Phase-resolved subspace images are matched to the dictionary to yield cine T₁, T₂, and M₀ maps. Images corresponding to the 2^nd and 3^rd singular values are used as bright and dark blood cine images.

Three healthy subjects were scanned at 1.5T (Sola, MAGNETOM Siemens) using cine MRF in a 10s breathhold. Three reconstructions were compared: low-rank subspace (LR), LR followed by nonrigid motion correction (LR-MOCO) as in (1), and the DIP method. Diastolic MRF, MOLLI, and T₂-prep bSSFP mapping and a standard cine were also acquired. Mean septal T₁ and T₂ and single-slice LVEF were compared among methods. In one subject, only the first 5s of cine MRF data was used to assess the effects of shortening the breathhold.

Results: Figure 2 shows DIP cine MRF maps and bright/dark blood cine images from one subject acquired in 10s. Figure 3 compares 10s and 5s scans using different reconstructions. For the 10s scan, improved image quality was observed with the DIP technique. For the 5s scan, LR and LR-MOCO could not recover meaningful maps, whereas DIP still performed well. Over all subjects, mean T₁ values were: cine MRF 1063ms (diastolic) / 1081ms (systolic), single-phase MRF 1062ms, MOLLI 1031ms. Mean T₂ values were: cine MRF 35.5ms (diastolic) / 36.9ms (systolic), single-phase MRF 35.2ms, T₂-bSSFP 49.6ms. Mean single-slice LVEF was 79.8% (cine MRF) versus 77.2% (standard).

Conclusion:

This study introduces a low-rank DIP reconstruction for cine MRF that may enable shorter exam times by providing simultaneous T₁, T₂, and M₀ maps and multi-contrast cine images. Since this technique yields true motion-resolved maps, it could be used to study changes in relaxation times over the cardiac cycle, which will be explored in future work.