Introduction

Low magnetic field (LF) MRI is currently gaining momentum as a complementary, adaptable approach to MRI diagnosis¹. However, the further impaired Signal-to-Noise Ratio (SNR), leading in turn to prolonged acquisition times, challenges its relevance at the clinical level. MRI sequences can be accelerated from undersampling strategies, but large undersampling factors can cause artefacts, such as blurring and severe aliasing when relying on the typical Fourier formalism. Recently, reconstructing an alias-free image using deep learning techniques has shown promising results^2,3. In this study, we leverage deep learning reconstruction to demonstrate the feasibility of highly undersampled (20% sampling) 3D LF MRI at 0.1 T (4.25 MHz). The model performance is evaluated on both retrospective and acquired, prospective 3D LF data.

Material and Method

A total of 19 sets of fully sampled, 3D spoiled gradient echo (GRE) MR images of the human wrist and hand were acquired at 0.1 T using a single Tx/Rx coil. Ten of the latter sets were acquired using the following parameters: matrix = 128 × 115 × 9, voxel size = 1.2 × 1.2 × 6.3 mm3, TE/TR = 7.2/31 ms, number of averages = 28 (acquisition time = 14 min 56 s). The other 9 sets were collected using heterogenous sets of sequence parameters in order to expand the size of our dataset, and overcome overfitting issues. In addition, data augmentation was applied using a 3D version of the data generator function from the Keras library⁴, capable of handling 3D image processing, including image flipping, shifting and rotation. k-space undersampling was performed on both the second and third phase encode directions (k_y and k_z) using a Gaussian mask where 20 % of the data was retained. Full and undersampled 3D k-spaces were resized in x and y-axes to reach a matrix size of 128 x 128. The full and undersampled 3D k-spaces were then inverse Fourier transformed, and normalized. It is important to mention that the reconstruction pipeline is a 2D approach (x-y axes), each “slice” being considered as one individual input image.
U-net⁵ was adopted as a convolutional neural network architecture and trained using the RMSProp optimizer with an adaptive learning rate. Network performance was assessed on 2 retrospective testing sets. Mean Squared Error (MSE) and Structure Similarity Index (SSIM) were chosen as evaluation metrics. Our model performance was finally evaluated on prospective undersampled images with the protocol and sampling scheme described above; the total acquisition time was 3 min 7 s.

Results

Fig.1 and 2 show respectively the reconstruction of retrospectively and prospectively undersampled data. As can be seen, the structures, the contrast, and the sharpness are mostly preserved in the reconstructed images. Quantitative results are summarized in table 1. The mean SSIM and MSE values support the effectiveness of U-net reconstruction in both retrospective and prospective experiments.

Discussion and conclusion

Despite the small and heterogenous LF MR training dataset, quantitative results on both retrospectively and prospectively accelerated acquisitions show that our approach is able to reconstruct 5-fold undersampled 3D LF MR data while maintaining anatomical structure, and preserving contrast. This could be a significant step towards clinical relevance of LF MRI. Future work will focus on improving the model performance by expanding the training dataset and further preserving high spatial frequencies, as well as using coils better suited for hand/wrist applications along with high performance imaging sequences.