CAMERA-NET: Cascade Multi-Level Wavelet neural network with data consistency for MRI Reconstruction

Motivation: The U-net architecture has experienced widespread application in deep learning-powered MR accelerated reconstruction. Nevertheless, the down-sampling and up-sampling operations embedded within the structure curtail the recovery capabilities of MR sparsely sampled signals. The U-net is widely used in deep learning-based MRI reconstruction. Its encoding-decoding component enlarges receptive field while the pooling and interpolation operation limits the ability to recover sparsely sampled MR signals. Goal(s): We propose a model that combines wavelet transformations and convolutional neural networks for MRI accelerated reconstruction, while incorporating data consistency constraints that represent a priori physical information. The wavelet transform and inverse wavelet transform are introduced to replace pooling and interpolation operations in order to maintain the spatial information of images during the encoding-decoding process within the neural network. Approach: We present a cascaded multi-level wavelet neural network with data consistency for accelerated MRI reconstruction framework, which we call CAMERA-NET. The CAMERA-NET introduces wavelet transform and inverse wavelet transform instead of pooling and unpooling to perform more accurate downsampling and upsampling. Simultaneously, we incorporate wavelet transform into the convolutional neural network, consequently refining its learning capability. Furthermore, to maintain a balance between high-frequency and low-frequency information, we suggest implementing the loss dropout technique to guide the neural network and preserve the important details. A cascaded multi-level wavelet neural network with data consistency, termed as CAMERA-Net, is presented for under-sampled MRI reconstruction. Results: Using the publicly available fastMRI knee datasets, it was found that upsampling and downsampling networks based on wavelet transformation produce better results compared to U-networks based on pooling and unpooling techniques. The inclusion of data consistency restrictions can additionally enhance the reconstruction performance. Reconstruction quality can be improved by increasing the number of iterations and optimising the loss dropout function. CAMERA-Net demonstrates significant enhancement in reconstructing quality with public fastMRI knee dataset. Impact: The improved reconstruction capabilities of CAMERA-Net have the potential to enhance precision and reliability when reconstructing under-sampled MRI data. This could result in more efficient clinical scans.