Volumetric Super-Resolution for Free-Running 3D Cardiac CINE Imaging on a 0.6T MRI

Background: Free-running 5D CMR enables cardiac and respiratory phase-resolved whole-heart imaging without ECG gating or breath-holding, allowing flexible reformatting in any view¹, and has been applied for biventricular assessment at 1.5T and extended to 0.6T^2,3. To keep scan times acceptable, aggressive undersampling is required, degrading image quality. We address this by reducing scan time with lower spatial resolution and fewer radial interleaves. Since circumferential spacing at the edge of k-space is maintained, fewer spokes suffice to avoid streaking, enabling low-resolution images that are enhanced post-acquisition with a 3D super-resolution model.

Methods: We employed a 3D generative AI model⁴ to enhance low-res images. Low-res data were generated with reduced 3D radial sampling (smaller k-space ball), lowering spatial resolution (Fig. 1A). Originally developed for 3D LGE imaging with Cartesian undersampling⁵, the model was adapted to 3D radial free-running (Fig. 1B). A fully convolutional 3D generator takes low-res input and outputs high-res volumes, using residual blocks to capture spatially coherent features. A 3D discriminator distinguishes real from generated data with adversarial training using voxel-wise, perceptual, adversarial, and k-space losses.

Training used data from 12 subjects scanned at 1.5T Ambition X and 0.6T Ambition X (Philips Healthcare, Best, The Netherlands) with FOV = 220 mm³, scan resolution 1.2-1.4 mm. Data were reconstructed into 4 respiratory and 14-21 cardiac phases, retrospectively undersampled 50%. Two subjects were held out for evaluation.

For prospective evaluation, 4 volunteers were scanned at 0.6T Philips system with a free-running 5D radial bSSFP sequence (TE/TR = 3.3/6.5 ms, flip angle 90°, superior-inferior readout every 15 TRs). Data were reconstructed into 4 respiratory and 20 cardiac phases. Each underwent high- and low-res scans: 8069 interleaves (1.38 mm, 13:08 min) and 3977 interleaves (2.4 mm, 6:31 min). Low-res volumes were enhanced with the trained model (Fig. 1C).

Results: Fig. 2A shows retrospectively simulated images of a subject where 3D-enhanced images resemble the high-resolution ground truth. Quantitative reference-free sharpness analysis⁶ (Fig. 2B) shows no significant difference between 3D-enhanced and high-res images (P=0.081), while low-res images show significant difference (P< 0.001).

Fig. 3A presents prospective scans, where low-res acquisitions (6:31 min) showed visual improvement after 3D enhancement compared with high-res scans (13:08 min). Fig. 3B summarizes a blind reader study across 4 subjects, where images were rated on a 4-point Likert scale for blurring (1 = none, 4 = severe). 3D-enhanced images scored 2.0 on average, compared with 2.75 for low-res and 2.25 for high-res.

Conclusion: We demonstrate that a 3D generative AI model enhanced low-res free-running CMR acquired in half the scan time. The 3D-enhanced images preserve quality and sharpness and enabling efficient, high-quality imaging at low field strength.

Figure 1: (A) Retrospective simulation of low-resolution training inputs: High-resolution 3D cardiac MRI datasets were retrospectively undersampled by 50% using a reduced 3D radial k-space mask. Each cardiac and respiratory phase was treated as a unique 3D volume for training the model. (B) A fully convolutional 3D generator network takes low-resolution volumetric input and outputs a high-resolution volume of the same size. The generator consists of multiple 3D residual blocks designed to learn spatially coherent features across all three dimensions. In addition to the generator, the model includes a 3D convolutional discriminator that distinguishes between real high-resolution images and those generated by the model. The network is trained adversarially, where the generator aims to produce images that are indistinguishable from ground truth, while the discriminator learns to classify real versus generated volumes. A combination of voxel-wise reconstruction loss, adversarial loss, perceptual loss, and Fourier loss (computed in k-space) is used to optimize the network during training, ensuring fidelity in both image and frequency domains. (C) Prospective inference pipeline: Low-resolution free-running CMR images acquired in reduced scan time were processed through the trained 3D super-resolution model to generate enhanced volumetric images for evaluation. In these acquisitions, both the spatial resolution and number of spokes were reduced, resulting in a smaller and less densely sampled koosh ball. Since the circumferential distance between spokes at the outer edge of k-space is preserved, fewer spokes are required to avoid streaking artifacts compared with the higher-resolution acquisition.
Figure 2: Retrospective evaluation of 3D-enhanced image quality compared to low- and high-resolution images. (A) Axial, sagittal, and coronal slices from a held-out subject show visual comparisons between high-resolution, low-resolution, and 3D-enhanced images. The 3D enhanced images reconstructed from 50% undersampled low-resolution data demonstrate close resemble to high-resolution images. (B) Quantitative sharpness analysis using a no-reference blur metric shows that 3D enhanced images have blur scores statistically indistinguishable from high-resolution images (P = 0.081), while low-resolution images show significantly degraded sharpness compared to high-resolution (P < 0.001).
Figure 3: Prospective evaluation of the super-resolution model on low-field 0.6T free-running cardiac MRI acquisitions. (A) Representative axial, sagittal and coronal slices from a volunteer scanned with both high-resolution and low-resolution protocols. Low-resolution scans were acquired in approximately half the scan time (6:31 min vs. 13:08 min). 3D-enhanced images generated from the low-resolution inputs show clear visual improvements, with structural details and myocardial borders more closely resembling the high-resolution reference. (B) Results of a blind reader study across four prospective subjects. Images were scored using a 4-point Likert scale (1 = no blurring, 4 = severe blurring). 3D-enhanced images achieved a mean score of 2.0, showing improvement over low-resolution (2.75) and closer to high-resolution (2.25).