Reconstruction of MRI from undersampled k-spaces of double-contrast volume acquisitions using deep neural networks

Magnetic Resonance Image (MRI) reconstruction from undersampled k-space using Deep Neural Networks (DNNs) has been extensively investigated. Undersampling accelerates the MRI acquisition process while addressing challenges such as motion artifact, signal decay, geometric distortion, and high Specific Absorption Rate (SAR). This article introduces a novel approach for the reconstruction of undersampled brain MRI, adopting a multi-slice and double-contrast (T1-weighted and T2-weighted) strategy. To tackle the undersampling challenge, we employ two complementary undersampling masks that alternate across different slices. This enables us to leverage the information from adjacent slices when estimating the unknown samples in the k-space domain. While both MR contrasts and all their slices are undersampled, we estimate the unknown samples in the k-space domain for each undersampled slice using known samples from adjacent slices and the corresponding slices of the other contrast. Subsequently, the estimated slices in the k-space domain for both contrasts are converted into the image domain. These converted images are then input into a U-net model, which further improves the quality of the estimated images. The U-net model takes both T1-weighted and T2-weighted images as input on two separate channels and predicts the T1-weighted image. The Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) of the reconstructed images from 30% of the k-space were 44.382 ± 2.139 and 0.987 ± 0.005, respectively. The proposed model achieved a higher PSNR with a lower variance using 30% and 40% under-sampling masks compared to the deep De-Aliasing Generative Adversarial Network (DAGAN) and self-Attention and Relative Average discriminator-based Generative Adversarial Network (SARA-GAN).