A Transfer-Learning Approach for Accelerated MRI using Deep Neural Networks

Neural network based architectures have recently been proposed for reconstruction of undersampled MR acquisitions. A deep network containing many free parameters is typically trained using a relatively large set of fully-sampled MRI data, and later used for on-line reconstruction of undersampled data. Ideally network performance should be optimized by drawing the training and testing data from the same domain. In practice, however, large datasets comprising hundreds of subjects scanned under a common protocol are rare. Here, we propose a transfer-learning approach to address the problem of data scarcity in training deep networks for accelerated MRI. The proposed approach trains neural networks using thousands of samples from a public dataset of natural images (ImageNet). The network is then fine-tuned using only few tens of MR images acquired in the testing domain (T1- or T2-weighted MRI). The ImageNet-trained network yields nearly identical reconstructions to networks trained directly in the testing domain using thousands of MR images, and it outperforms conventional compressed sensing reconstructions in terms of image quality. The proposed approach might facilitate the use of neural networks for MRI reconstruction without the need for collection of extensive imaging datasets.