Nonrigid MOtion COrrected DEep learning-based rapid Spiral Image REconstruction (MOCO-DESIRE) for free breathing, high-resolution subspace navigator guided first-pass myocardial perfusion imaging at 3T

Motion-corrected high-resolution whole-heart first-pass myocardial perfusion imaging during free breathing can be achieved by integrating a fixed-angle subspace navigator-guided spiral acquisition with a multi-step MOCO incorporated subspace reconstruction¹. This pipeline typically begins with an auxiliary reconstruction with high motion fidelity to estimate the deformation field, followed by a reconstruction which incorporates rigid-nonrigid motion into the forward model of subspace reconstruction^1-3.  These MOCO incorporated reconstruction reduce motion blurring and enhance edge sharpness. However, the required multiple iterative reconstructions and inverse deformation field estimation can be time-consuming, making rapid on-line reconstruction challenging. Deep learning-based reconstructions have emerged as effective solutions for fast cardiac imaging, enabling reconstruction within seconds^4-6. Notably, a DESIRE network based on U-Net has shown excellent performance for spiral perfusion imaging⁶. The goal of this work is to adapt the DESIRE network for our MOCO-subspace reconstruction, aiming to reduce the reconstruction time from 1-hour/slice to a few seconds.



Methods:

Free-breathing perfusion images were acquired from 17 subjects undergoing clinical CMR studies using a 3T Siemens scanner. The imaging protocol employed seven golden-angle rotating spiral interleaves combined with a fixed-angle navigator trajectory. Acquisition parameters included 6 slices, resolution = 1.3×1.3 mm², slice thickness = 10 mm, TR = 6ms, temporal footprint = 90ms, saturation time = 60ms, and FA = 35°.

The MOCO-DESIRE pipeline is shown in Figure1.  13 datasets were used for training, 1 dataset for validation and 3 datasets for testing. The inputs were normalized by their maximum absolute value. The reconstructed images were visually assessed by a cardiologist using a 5-point scale (1 = poor to 5 = excellent). Temporal fidelity was evaluated in the ventricular with MOCO-SENSE as the reference.



Results:

Figure 2 shows the predicted images with motion from the 2D-UNet and predicted images with MOCO from the 3D-UNet. The 2D-UNet effectively reduces undersampling artifacts while maintaining high motion and temporal fidelity. Meanwhile, the 3D-UNet substantially mitigates aliasing, preserves fine tissue structures, and minimizes blurring.

Figure 3 presents the good alignment and high temporal fidelity achieved using the MOCO-DESIRE pipeline. The deformation field and the temporal basis extracted from the 2D-UNet predictions exhibit high fidelity, serving as a foundation for the 3D-UNet’s high-temporal fidelity output. 

The image quality scores for the MOCO-DESIRE and MOCO-subspace were similar (3.5 vs 3.5 respectively, p=0.33), and both were substantially higher than that of MOCO-SENSE (2.3, p< 0.001).



Conclusion:

By integrating a 2D-UNet for the high-fidelity motion image reconstruction with a 3D-UNet for MOCO image reconstruction, the 1-hour/slice reconstruction can be substantially shortened, and the MOCO-DESIRE pipeline should cost a time of 45 seconds enabling on-line reconstruction.  The MOCO-DESIRE results exhibit high quality, sharpness, and alignment comparable to those achieved with MOCO-subspace for 1.3x1.3 mm² free-breathing cardiac perfusion imaging.


Figure 1 illustrates the flowchart of the proposed MOCO-DESIRE pipeline. Initially, a 2D-UNet with a depth of 3 and 16 initial kernels (D3K16) was used to reconstruct the images with motion. The 2D-UNet allows for high motion fidelity and faster training. In this step, temporal basis-informed NUFFT images served as network input, with auxiliary subspace reconstruction images as the ground truth. The deformation field and temporal basis were then extracted from the 2D-UNet reconstructed images. Subsequently, a 3D-UNet with more filters (D3K16) and 3D convolutions was applied for the final MOCO reconstruction. This deeper network with more filters could better leverage spatiotemporal information, capturing finer structural details. At this stage, the temporal basis and deformation informed NUFFT images were used as inputs, and the MOCO-incorporated subspace images served as the ground truth.
scmr2025-figure1.pdf
Figure 2 presents four frames with dynamically varying contrast across different ventricles in the temporal basis-informed NUFFT inputs, the U-Net predictions, the ground truths (GT), and the difference maps between the predictions and the GTs for both motion images reconstructed by the 2D-UNet and MOCO images reconstructed by the 3D-UNet, from two subjects. The top four rows display the 2D-UNet reconstructions of images with respiratory motion, while the bottom four rows show the 3D-UNet reconstructions of MOCO images. Blue arrows indicate areas where aliasing artifacts are substantially reduced by the U-Net networks, and yellow arrows highlight regions where fine tissue structures are well preserved, with improved sharpness of subtle structures in the 3D-UNet reconstructions.
scmr2025-figure2.pdf
Figure 3A illustrates the good alignment of free-breathing respiratory images using the MOCO-DESIRE pipeline by comparing different time frames and 1D-t profiles of reconstructions with and without MOCO. Red lines serve as baselines to compare motion across frames, with the corresponding 1D-t profiles shown at the end of each row. The white line in the top-right image indicates the position used for the 1D-t profile. The 2D-UNet output demonstrates a similar respiratory motion pattern to the motion ground truth. All frames from the 3D-UNet are well-registered to the frame with the highest contrast in the two ventricles. Figure 3B shows the averaged temporal curves in the left and right ventricles. The black, blue, and red lines represent the curves for MOCO-SENSE, MOCO-GT, and MOCO-DESIRE reconstructions, respectively. The masks of the ventricles are shown on the top left of the plots. A high level of consistency between the MOCO-DESIRE and MOCO-SENSE curves, as well as between MOCO-DESIRE and MOCO-GT can be observed.
scmr2025-figure3.pdf