Analysis of Deep Complex-Valued Convolutional Neural Networks for MRI\n Reconstruction

Abstract

Many real-world signal sources are complex-valued, having real and imaginary\ncomponents. However, the vast majority of existing deep learning platforms and\nnetwork architectures do not support the use of complex-valued data. MRI data\nis inherently complex-valued, so existing approaches discard the richer\nalgebraic structure of the complex data. In this work, we investigate\nend-to-end complex-valued convolutional neural networks - specifically, for\nimage reconstruction in lieu of two-channel real-valued networks. We apply this\nto magnetic resonance imaging reconstruction for the purpose of accelerating\nscan times and determine the performance of various promising complex-valued\nactivation functions. We find that complex-valued CNNs with complex-valued\nconvolutions provide superior reconstructions compared to real-valued\nconvolutions with the same number of trainable parameters, over a variety of\nnetwork architectures and datasets.\n