Time-dependent Deep Image Prior for TR-resolved cine MRI at 0.55T

Background: Conventional 2D cardiac cine MRI is typically reconstructed with a resolution of ~30 frames by combining data acquired over 10-15 heartbeats¹. To avoid averaging effects over different cardiac cycles, MC-CINE² demonstrated the feasibility of single-heartbeat cine imaging at 1.5T using radial acquisitions with motion correction and patch-based regularization. However, this approach has limited temporal resolution due to computational complexity scaling quadratically with the number of temporal bins. Self-supervised methods have shown promise in highly undersampled MRI at conventional fields, while also enabling reconstructions at arbitrary temporal resolution without extra computational cost³. However, low-field cardiac MRI at 0.55T presents further challenges due to reduced SNR, longer TR, fewer spokes per heartbeat, and more aggressive undersampling⁴. Here, we investigate the potential of Time-Dependent Deep Image Prior (TD-DIP)³ for TR-resolved radial cardiac cine MRI reconstruction at 0.55T without temporal binning.

Methods: Radial cine acquisitions were performed with a bSSFP readout and tiny golden-angle (23.6°) at 0.55T (MAGNETOM Free.Max, Siemens Healthineers) using a 15-channel cardiac coil. Reconstruction was performed using TD-DIP, incorporating cardiac cycle and phase (obtained from ECG signals from the scanner) through a helix encoding (Fig 1). This enabled TR-resolved reconstructions at the acquired temporal resolution while distinguishing between individual cardiac cycles. A fully sampled 15s reference acquisition was retrospectively undersampled to reconstruct the first N cardiac cycles, with N∈[1, …, 8] and N=N_max (fully sampled) for comparison.
Experiments were conducted in a mid-ventricular short-axis slice for 6 healthy subjects. Imaging parameters were: FOV=256×256 mm², slice thickness=10mm, in-plane resolution=2×2mm², TE/TR=2.85/5.7ms, flip angle=78°. TD-DIP reconstructions were trained for (2 + number of heartbeats/2)×1000 iterations with a learning rate of 2×10^-3.

Results: Comparable depiction of cardiac dynamics to the fully sampled reference was achieved with as few as 2-3 heartbeats (Fig 2.a). For all subjects, image quality metrics improve when adding more cardiac cycles to the reconstruction (Fig 2.b). TR-resolved TD-DIP preserved more temporal details than the binned reference, while TR-resolved reconstructions from only 2 heartbeats (without cardiac cycle averaging) provided comparable quality against full-data TR-resolved reconstructions (Fig 3).

Conclusion: TR-resolved TD-DIP enabled high-quality cardiac cine reconstructions at 0.55T, revealing subtle temporal dynamics that were absent in binned cine references. This is consistent with previous findings^2,5, which reported improved sharpness when avoiding averaging data from multiple cardiac cycles. Importantly, the ability to reconstruct with only 2-3 heartbeats while maintaining image quality highlights the potential of TR-resolved TD-DIP for rapid cardiac cine imaging at low field.

TD-DIP with Helix encoding is trained in different number of cardiac cycles. The number of twists in the helix corresponds to the number of heartbeats used in the reconstruction. A representative temporal profile for the TR-resolved (1 radial spoke per cardiac phase) cine image over the entire acquistion (15s) is shown at the bottom.
A) Reconstructions for various cardiac cycles for two subjects (fast and medium heart rate) in comparison to fully sampled binned reference (first column). TR-resolved DIP reconstructions are TR-resolved, only the first heartbeat is shown. Red arrows indicate blurriness in the binned reconstruction that appears sharper in the TR-TD-DIP reconstruction B) PSNR for TR-TD-DIP for all subjects using from 1 to 8 cardiac cycles, calculated at first heartbeat, against TR-TD-DIP reference trained with all cardiac cycles.
Vertical and horizontal temporal profiles for all subjects. Binned Fully Sampled (FS) reconstruction (30 cardiac phases binned over all cardiac cycles), TR-TD-DIP Fully Sampled, using all the acquired data but TR-resolved, and TR-TD-DIP using just the first two cardiac cycles and TR-resolved. Reconstructions are shown at first cycle. Subjects are sorted from left to right in ascending order of the mean number of spokes per heartbeat (hb), i.e. from most challenging to the easiest.