Self-supervised Super-resolution based on Implicit Neural Representation for 3D whole-heart CMRA at 0.55T

Background: Free-breathing 3D whole-heart cardiac MRA (CMRA) has been recently demonstrated at 0.55T in about 6 min scan time with image-navigator (iNAV) based non-rigid motion corrected reconstruction [1]. However, scan time remains lengthy for clinical settings. Deep learning super-resolution (SR) approaches can be exploited to further accelerate the acquisition by acquiring a low resolution (LR) volume and learning a high-resolution (HR) representation of the image [4]. However, most of these approaches are supervised and require large amount of data for training, which is limited at low-field. Recently, a self-supervised Implicit Neural Representation (INR) that enables SR of arbitrary scales have been proposed [3]. Here we propose to extend this approach to enable 4x downsampled CMRA at 0.55T.

Methods: iNAV-based CMRA at 0.55T with 1.5 mm³ resolution (HR) was employed (208x208x88 matrix size). LR images (104x104×44 for 4x) were retrospectively downsampled in k-space (phase-encoding plane: RL, AP) with a factor of 4× (2× per direction).

Training was performed per phase-encoding plane. For 4× upsampling, paired LR (104×44) and auxiliary LR2 (52×22) inputs were used for training (Fig.1a). These pairs were fed to the encoder to extract feature maps, with the Mean Absolute Error (MAE) loss computed in image space. The decoder predicted voxel intensities at any spatial coordinate within the predicted SR volume.

Inference for 4x was from acquired LR (104×44) to SR (208x88) dimension (Fig.1b). The proposed approach was compared with Bicubic Interpolation using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) as evaluation metrics.

Results: SR results for 4x upsampling with the proposed self-supervised INR approach are shown in Fig.2 for a representative subject. Comparison against bicubic interpolation and the HR reference are shown for the axial plane (Fig.2a). Results for the proposed approach for the other views is shown in Fig2.b in comparison to the HR reference. Excellent visual image quality is observed with the proposed approach for all views, which is consistent with the quantitative metrics (Table 1).

Conclusion: his work presents a novel self-supervised INR SR framework to reconstruct 3D CMR images at 0.55T without any external training data. Results suggests than a 4x scan time reduction could be possible with the proposed approach, without affecting image quality. Validation in prospectively downsampled data is now required. Future work includes incorporating data consistency and comparison against other SR frameworks in a large cohort of healthy subjects and patients with cardiovascular disease.

PSNR and SSIM quantitative results of our self-supervised INR method and Bicubic

Down sampling	Our model		Bicubic Interpolation
	SSIM	PSNR	SSIM	PSNR
4x	0.8980	31.96	0.7663	25.83

Figure 1: a) Training was performed per phase-encoding plane (axial). For 4× upsampling, paired LR (104×44) and auxiliary LR2 (52×22) inputs were used for training. The INR SR model, consists of an RDN encoder and an MLP decoder. b) Inference for 4x upsampling was from acquired LR (104×44) to SR (208x88) dimension.
Figure 2: a) SR results for 4x with the proposed self-supervised INR approach in comparison to bicubic interpolation and the reference HR in the axial plane. 10% error images are also included for both bicubic interpolation and our INR SR method. b) Results for the proposed approach for 4x for coronal and sagittal views in comparison to HR reference.