Learned Multi-level Wavelet for Fast MRI Reconstruction

Background: Magnetic Resonance Imaging (MRI) is a widely applied medical imaging technology for various clinical applications. However, one major drawback is the long scan time, which results in significant artifacts due to patient’s voluntary and non-voluntary movement. This work aimed to expedite the scan time by reconstructing diagnostic MRI images from an undersampled k-space.

Methods: To solve this ill-posedness, wavelet sparse transform was introduced to provide a successful model for optimizing and regularizing the inverse problem. It offers a complete representation of the image and accurate analysis at different spatial orientations. To that end, the implicit sparsity in MRI images is employed to undersampled k-space. In this study, we exploited MLTL, which is an iterative multi-level wavelet algorithm, to reduce MRI acquisition time and improve reconstruction. MRI brain experiment was carried out using multi-channels coil model to evaluate the parallel imaging acquisition. To evaluate the MLTL algorithm with this multi-channel coil model, different undersampling implementations were used. The acceleration factors were set to 5 and 10. The performance of the proposed algorithm is demonstrated and compared with that of the traditional reconstruction method (IFFT) and another competitive algorithm (FISTA).

Results: The final numerical simulation showed an outperformance of the proposed MLTL algorithm in terms of the accuracy of the reconstructed image and acceleration of the reconstruction time. MLTL reconstructed high-resolution images from the undersampled measurements. It maintained the details and recovered the structures of the image more accurately. Meanwhile, IFFT showed very noisy images and FISTA provided blurry images, both of which cannot be used clinically. The MSE and PSNR were calculated for the three methods, and the MLTL-reconstructed images showed less error and the best image resolution.

Conclusions: MLTL can be successfully applied to undersampled k-space reconstruction, which has the potential to accelerate clinical acquisition