Multi-slice and multi-contrast brain MRI reconstruction from undersampled k-space using deep neural networks

Background and aims: Deep Neural Networks have been widely employed for MRI (MagneticResonance Imaging) reconstruction from undersampled k-space, and their performance has suppressedthe performance of conventional algorithms and methods. Undersampled k-space dataacquisition can enhance the acquisition speed while addressing issues such as motion artifact,blurring, high SAR.Method: We propose a multi-slice and multi-contrast (T۱-weighted and T۲-weighted) approachfor reconstruction of the undersampled brain MRI. The proposed network relies on the mutualinformation between adjacent slices and various contrasts to obtain better results. The undersamplingmasks are designed in a complementary manner to reduce the redundancy of the acquiredsamples. The unacquired samples of each MRI contrast are first estimated from known samples ofadjacent slices and the other contrast in the k-space domain (using K-Space-T۲ and K-Space-T۱models) then a U-net model is used for further improvement of the estimated images, which takesboth T۱-weighted and T۲-weighted images as input and predicts the T۱-weighted image. TheBRATS ۲۰۱۵ dataset is used for training and testing of the network.Results: The PSNR and SSIM of the reconstructed images using ۳۰% of the k-space sampleswere ۴۴.۳۸۲±۲.۱۳۹ and ۰.۹۸۷±۰.۰۰۵ respectively. We were able to achieve a better PSNR withlower variance for ۳۰% and ۴۰% under-sampling masks when compared to DAGANand SARA-GAN.Conclusion: Using mutual information between adjacent slices and complementary contrast canenhance the quality of MRI reconstruction. The empirical results shows that our proposed methodhas lower variance and subsequently worst-case situations are less severe. From the providedreconstructed images, it is evident that the error map is homogenous over the whole brain andthe tumors are reconstructed as good as other tissues. It is important to note that Complementarymask design makes T۱-weighted and T۲-weighted series interdependent, which may introducenew artifacts and difficulties in practice. Also, the model assumes that the slice distance and imagingplane for each series are small and identical.

کلیدواژه ها:

Brain MRI Reconstruction ، Deep Neural Networks ، Multi-slice ، multi-contrast ، Undersampled K-space