Motion-Augmented End-to-End Nonrigid MOCO-DESIRE for Free-Breathing Spiral First-Pass Myocardial Perfusion Imaging at 3T with Limited Training Data

Background: Whole-heart first-pass myocardial perfusion imaging during free breathing enables high-resolution assessment but is vulnerable to respiratory motion, which degrades spatiotemporal fidelity¹. Conventional MOCO reconstructions mitigate motion but require iterative steps taking up to an hour per slice². Deep learning offers a faster alternative: DESIRE uses a U-Net to suppress motion artifacts and preserve temporal fidelity³. MOCO-DESIRE extends this with a VoxelMorph-based motion estimator⁴ and 3D U-Net reconstruction, enabling end-to-end training that reduces reconstruction times to < 5 s while improving structural fidelity⁵.

Sharp myocardial perfusion images are critical, yet residual respiratory motion can blur reconstructions and impair deep learning performance. Motion augmentation improves robustness by exposing the network to realistic deformations⁶, while fixed-frame or fused-template references often fail due to contrast mismatch, and PCA-based motion suppression can oversmooth fine structures⁷. We incorporate multistage motion augmentation and sharp PCA-synthesized references into MOCO-DESIRE to enhance performance with limited datasets.

Methods: Perfusion CMR data were acquired from 19 patients on a 3T Siemens scanner using seven golden-angle rotating spiral interleaves with a fixed-angle navigator, yielding six 10-mm slices at 1.3 × 1.3 mm² resolution, TR = 6 ms, and 90-ms temporal footprint. We developed a motion-augmented end-to-end MOCO-DESIRE pipeline integrating a 2D U-Net–based VoxelMorph for nonrigid motion estimation with a 3D U-Net–based DESIRE for image reconstruction. MOCO-subspace reconstructions served as ground truth. To improve robustness, only moving images were augmented via RandAffineD (rigid shifts) and Rand2DElasticD (localized deformations) in a three-stage curriculum (warm → ramp → full), progressively increasing motion severity. During testing, sharp PCA-synthesized references were used since motion-free images are unavailable. PCA reconstructions were denoised (~90% variance) and sharpened with Sobel-weighted unsharp masking to provide contrast-consistent inputs. Training/validation/testing split was 15/1/3.

Results: Figure 2 shows motion-augmented training enables the network to capture more cardiac motion, improving U-Net reconstruction sharpness. Motion-augmented reconstructions were significantly sharper on temporally projected images (0.15 ± 0.13) than non-augmented MOCO-DESIRE (0.13 ± 0.11, P < 0.001). Figure 3 shows Sharp PCA references for VoxelMorph yield sharper reconstructions than conventional PCA.

Conclusion: Multistage motion augmentation in the end-to-end MOCO-DESIRE framework enables robust correction of respiratory motion, even with limited training data. PCA-based edge-sharpened references preserve fine anatomical details, avoiding oversmoothing. Together, these strategies allow rapid, high-quality 3D DESIRE reconstructions from free-breathing spiral perfusion data, demonstrating strong potential for real-time clinical use.

Figure 1. Overview of the motion-augmented MOCO-DESIRE pipeline. The end-to-end network integrates a 2D U-Net–based VoxelMorph for nonrigid motion estimation with a 3D U-Net–based DESIRE for image reconstruction. During training, MOCO-subspace datasets serve as ground truth, and NUFFT images are generated from a navigator-derived temporal basis. Multistage motion augmentation introduces progressively larger translations, rotations, and elastic deformations, while PCA-synthesized, edge-sharpened images provide sharp references for registration during testing. VoxelMorph applies the estimated deformation to both real and imaginary components of the moving images, which are then reshaped into 3D format for DESIRE reconstruction. The network is trained using a combination of MSE of reconstruction, NCC of registration, and a deformation-smoothness regularization term.
Figure 2. Two representative frames from the MOCO-DESIRE reconstructions and the corresponding VoxelMorph-predicted deformation fields with and without motion-augmented training. The first row shows pixelwise deformation field magnitudes, where larger values indicate greater estimated motion. Red arrows highlight regions in the heart where motion-augmented training enabled VoxelMorph to capture more motion, leading to sharper reconstructions in the second row.
Figure 3. Comparison of free-breathing respiratory images across time frames and corresponding 1D-t profiles using MOCO-DESIRE with conventional PCA and proposed Sharp PCA as VoxelMorph registration references. Blue lines indicate baselines for cross-frame comparison, with 1D-t profiles at the end of each row. Red arrows highlight regions of high sharpness in the Sharp PCA–referenced reconstruction. The blue arrow shows strong artifact suppression, while the green arrow indicates slight residual motion not fully corrected by conventional PCA-guided MOCO-DESIRE or MOCO-subspace, but corrected by proposed Sharp PCA.