A Fast Deep Image Prior Reconstruction for Cardiac MR Fingerprinting Using Meta-Learning Acceleration

Background: High-quality cardiac MR Fingerprinting (MRF) maps can be obtained using a Deep Image Prior (DIP-MRF) reconstruction, which combines a low-rank model with zero-shot deep learning¹. However, this technique requires long reconstruction times ( >40min on a GPU) since training is performed in a scan-specific manner. This work introduces an approach to shorten the reconstruction time to 5min per slice using meta-learning with scan-specific fine-tuning.

Methods: DIP-MRF employs: (1)a U-Net (Image Reconstruction Network or IRN) that outputs subspace images from a fixed noise input tensor; (2)a convolutional network (Coil Estimation Network or CEN) that refines sensitivity maps from ESPIRiT², and (3)a fully-connected network (Parameter Estimation Network or PEN) that estimates T₁, T₂, and M₀ maps³. These networks are trained de novo to enforce consistency with k-space data from a single scan. Here, the following approach is proposed to accelerate DIP-MRF (Figure 1).

Step0: The IRN and CEN are pre-trained using scans from multiple subjects (using the same input tensor for the IRN) by minimizing the MSE relative to target images and sensitivity maps from the original (slow) DIP-MRF. Rather than simple pre-training, this step uses REPTILE⁴, a meta-learning algorithm that allows pre-trained networks adapt to new datasets in fewer steps

Step1: The pre-trained IRN and CEN are fine-tuned for a specific scan. To accelerate this process, the loss does not include a forward physics model but instead minimizes the MSE between network-generated and zero-filled (undersampled) images in the subspace

Step2: Fine-tuning is then performed using a loss incorporating the forward physics model to enforce consistency with acquired k-space data

Step3: The PEN is trained in a self-supervised manner to estimate quantitative maps

Note that the original DIP-MRF method is equivalent to only performing Steps 2-3. Data were collected at 1.5T (MAGNETOM Sola, Siemens) using a 2D breathheld cardiac MRF sequence in 28 healthy subjects at a mid-short-axis slice⁵. Meta-learning was performed using data from 20 subjects; the remaining 8 were used for testing, with 1000 epochs for Step1 and 10 epochs for Step2. For comparison, the original DIP-MRF was performed with 300 epochs.

Results: MRF maps after each stage of the accelerated reconstruction are shown in Figure 2. After Step0, outlines of anatomical structures can be observed. After Step1, higher-quality maps are obtained, although substantial aliasing artifacts remain. After Step2, these artifacts are largely suppressed, yielding maps comparable to the original DIP-MRF technique in only 5min (compared to 40min). Subtle spatial blurring is observed relative to the original method, which can be mitigated by increasing the epochs in Step2 to 100 (Figure 3) at the expense of a longer reconstruction time (18min).

Conclusion: This work introduces a meta-learning strategy to accelerate DIP-MRF, which could enable rapid, inline deep learning reconstruction of cardiac MRF maps.

Figure 1: Overview of the proposed meta-learning pipeline to accelerate DIP-MRF reconstruction for cardiac MRF. (A) Step 0: Meta learning is performed using 20 datasets previously reconstructed using the original (slow) DIP-MRF method to pre-train the IRN and CEN. (B) Step 1: The IRN and CEN are adapted to data from a new subject using a simplified loss function that does not enforce data consistency. (C) Step 2: Fine-tuning is performed (for a small number of epochs) using a loss including the forward model to enforce data consistency. (D) Step 3: The PEN is trained in a zero-shot manner to estimate quantitative maps. Reconstruction results of T1, T2, and M0 maps at different stages of the proposed accelerated DIP-MRF compared with the original DIP-MRF.
Figure 2: Reconstruction results after different stages of the accelerated DIP-MRF pipeline. (A) Maps are shown after meta-learning (Step 0), followed immediately by Step 3 (tissue property estimation). (B) Maps are presented after meta-learning and coarse fine-tuning (Step 1) for a specific subject. (C) Maps are displayed after meta-learning, coarse fine-tuning, and fine-tuning with the forward model (Step 2). The total reconstruction time (accounting for Steps 1-3) is ~5 min. (D) Results from the original (slow) DIP-MRF reconstruction are shown for comparison, with a reconstruction time of 42 min.
Figure 3: Comparison of accelerated DIP-MRF and original DIP-MRF reconstructions in two representative healthy subjects. Maps using the accelerated reconstruction are shown using both 10 epochs and 100 epochs for Step 2. T1 and T2 values measured in the left ventricular septum are reported as mean ± standard deviation.