Deep learning based image enhancement for dynamic non-Cartesian MRI: Application to “silent“ fMRI

Radial based non-Cartesian sequences may be used for silent functional MRI examinations particularly in settings where scanner noise could pose issues. However, to achieve reasonable temporal resolution, under-sampled 3D radial k-space commonly results in reduced image quality. In recent years, deep learning models for improving image quality have emerged. In this study, we investigate the applicability of deep learning image enhancement methods with a focus on preserving dynamic temporal signal changes. By utilizing high-resolution resting-state fMRI datasets from the Human Connectome Project (HCP) foundation, a ground-truth training set was constructed. The k-space trajectory coordinates of a so-called silent 'Looping Star' fMRI sequence was used to simulate non-Cartesian MRI data from the HCP datasets. Subsequently, these sparse resampled k-space were reconstructed, thereby generating pairs of simulated 'Looping Star' images and ground truth HCP images. The dataset served as the basis for training both 2D-UNet and 3D-UNet deep learning models for image enhancement. A comparative analysis was conducted, and the superior model was further fine-tuned. Evaluation of the final model's performance included standard image quality metrics as well as resting-state fMRI (rs-fMRI) analysis in the time-domain. The 3D-UNet outperformed the 2D-UNet in the image enhancement task, resulting in a significant reduction in error between the network input and the ground truth. Specifically, the 3D-UNet achieved a 97 % reduction in the mean square error between the simulated Looping Star input and the HCP ground truth in the pre-processed dataset. Moreover, the 3D-UNet successfully preserved voxel variations, observed as the correlated activity in the posterior cingulate cortex (PCC) during rs-fMRI analysis while simultaneously mitigating noise in the time-series images. In summary, image quality was improved and artifacts were effectively eliminated through the application of both 2D and 3D deep learning approaches. Comparative analysis of the networks indicated that the use of 3D convolutions is more advantageous than employing a deeper network with 2D convolutions, particularly in scenarios involving global artifacts. Furthermore by demonstrating that the trained neural network successfully preserved temporal characteristics in the BOLD signals, the results suggest applicability in fMRI studies.