Comparison of Neural Network Architectures for Physics-Driven Deep Learning MRI Reconstruction

Machine learning techniques have recently received interest as a means of improving MRI reconstruction. Conventionally, ill-conditioned reconstruction problems are solved using iterative optimization algorithms that alternate between applying data consistency and a proximal operator based on a regularizer. This iterative procedure can also be unrolled for a finite number of iterations to generate a feed-forward model. In physics-driven machine learning approaches, the known forward encoding model is used for enforcing data consistency in an unrolled iterative regularized least squares reconstruction. A neural network, which may or may not share weights across different unrolled iterations, is used as the regularizer prior. In this study, we aim to compare several neural network architectures, namely U-Net, ResNet and DenseNet for such physics-driven reconstruction. The performance of these architectures are evaluated on the publicly available fastMRI knee database. Comparisons are made for uniform and random undersampling masks. The results indicate that a DenseNet regularization unit performs as well as the other strategies for both uniform and random undersampling patterns, even though it has considerably fewer trainable parameters.