PARALLEL MAGNETIC RESONANCE IMAGING USING NEURAL NETWORKS Neelam Sinha Manojkumar Saranathan K. R. Ramakrishnan Department of Electrical Engineering, Indian Institute of Science, Bangalore, India S. Suresh School of EEE, Nanyang Technological University, Singapore ABSTRACT Magnetic resonance imaging of dynamic events such as cognitive tasks in the brain, requires high spatial and temporal resolution. In order to increase the resolution in both domains simultaneously, par- allel imaging schemes have been in existence, where multiple re- ceiver coils are used, each of which needs to acquire only a frac- tion of the total available signal. In our approach, we regularly un- dersample the signal at each of the receiver coils and the resulting aliased coil images are combined (unaliased) using the neural net- work framework. Data acquisition follows a variable-density sam- pling scheme, where lower frequencies are densely sampled, and the remaining signal is sparsely sampled. The low resolution images ob- tained using the densely sampled low frequencies are used to train the neural network. Reconstruction of the image is carried out by feeding the high-resolution aliased images to the trained network. The proposed approach has been applied to phantom as well as real brain MRI data sets, and results have been compared with the stan- dard existing parallel imaging techniques. The proposed approach is found to perform better than the standard existing techniques. Index Terms— Parallel Magnetic Resonance Imaging, under- sampling, unaliasing, neural networks 1. INTRODUCTION Magnetic Resonance Imaging (MRI) is a very popular medical imag- ing modality, due to its non-invasive nature and excellent soft-tissue contrast. MR signal acquisition occurs in the spatial-frequency do- main (called k-space), of the object being scanned. The k-space samples are used to construct the image of the object. Spatial lo- calization is achieved using magnetic field gradients applied in all three dimensions, which modulate the precession frequency of the protons as a function of space. The gradient along Z is called, slice selection, which determines the cross-section of interest. The gradi- ents along Y and X, called phase and frequency encoding gradients, determine in-plane resolution and field of view. To create an MR im- age, we need to sample the two-dimensional k-space (Δkx, Δky ). Here, Δkx = γ 2π GxΔt and Δky = γ 2π Gy τpe, where γ is a constant associated with hydrogen protons, Gx is the frequency encoding gra- dient amplitude, Δt is the sampling period, Gy is the phase encoding gradient step size, and τpe is the phase encoding gradient duration. In order to obtain an image with good spatial resolution, we need to acquire data farther out in k-space, implying larger number of phase encoding steps. However, an inherent limitation in MR imag- ing is that only one point in k-space can be acquired at any given instant of time. Hence, acquisition of more points would imply loss of resolution in time. Alternately, increasing Δky can reduce the number of phase encoding steps but at the cost of aliasing. This is the classic trade-off between spatial and temporal resolution in MRI, (a) (b) Fig. 1. Effect of downsampling (a) True image (b) Aliased image obtained by downsampling by 2 which is typically addressed using partial data reconstruction or par- allel imaging techniques. Parallel imaging involves use of multiple receiver coils. Here, acceleration is achieved by regular undersam- pling and the resulting aliasing resolved by making use of redundant information collected from multiple parallel receivers. The final un- aliased reconstructed images can be obtained by combining the sig- nals either in image domain or in k-space. Existing techniques differ in aspects such as the domain they work in, the assumptions they make, and the nature of errors that they generate. 2. PARALLEL IMAGING Parallel imaging was designed as a method to reduce the number of phase-encoding steps, the most time-expensive factor in MR Imag- ing. Here, multiple receiver coils are used in order to accelerate imaging. Each receiver coil is characterized by its spatial sensi- tivity function, which conveys information about the relative posi- tion of the origin of the received signal. Each coil provides the coil-weighted version of the image, all of which eventually can be combined to obtain the image reconstruction. It is well-established that if each of the receiver coils could acquire the entire k-space, then the best estimate of the true k-space would be the “sum of squares” (SoS). However, when the k-space at each of the receiver coils is sparsely sampled, then we need to devise ways to combine the acquired signals, in order to reconstruct the image. Methods like SENSE [1], SMASH [2], PILS [3], GRAPPA [4] are the known stan- dard techniques used in parallel imaging. 2.1. Current techniques SENSE combines the acquired signal in the image domain. Here, coil sensitivity information is used to combine the coil-weighted aliased images. Let us assume i1 and i2 to be the true intensities at the pixels shown in Fig.1(a). Let the coil sensitivities at those points be c11, c12 and c21,c22 for coils 1 and 2 respectively. The resulting intensity at the pixel marked in Fig.1(a), for coil 1 is say III - 149 1-4244-1437-7/07/$20.00 ©2007 IEEE ICIP 2007