Item

Abstract

Image reconstruction is the first post-processing step in magnetic resonance (MR) imaging and determines the base quality for further processing. Modern techniques driven by sparsity constraints lead to better image quality and more options to accelerate the acquisition compared to methods strictly based on analytical models of the imaging process. Deep Decoder (DD) [1], a promising approach of this kind, tunes a convolutional neural network (CNN) as an adaptive function transforming a fixed noise vector to the image of interest. The CNN represents an implicit prior on the structure of clean images leading the network to create essentially a denoised or artifact reduced version of the target image. DD has been successfully applied to MR data with Poisson-sampling [2]. Looking for improved image reconstruction of our own multi-parametric mapping (MPM) MR data [3] acquired in sub-sampled cartesian grid pattern, we want to evaluate the applicability of DD. Following [2] the original DD was modified to process axial slices of MPM data of a human head. The subsampled k-space was zero filled, Fourier transformed, sliced in readout direction and zero padded to 256x256 voxels before processing. Final magnitude images were generated by root-sum-of-squares combination across channels. For comparison different number of filters per layer (128 to 512) and iterations (10k to 100k) were tested. Qualitative inspection showed that models using less than 384 filters reconstructed only very blurry images. Output quality generally increased with number of iterations. All reconstructed images retained a significant amount of SENSE ghosting, specifically for high contrast regions like the skull, rendering the results unusable. Since the DD relies on the affinity of CNNs for structures over random noise, this is not surprising and difficult to counteract. With the cartesian sampling pattern given, DD can in consequence only be used to target data without structured artifacts. In conclusion the DD as used in [2] is not directly applicable to data sampled in cartesian fashion. Future research could instead try to include DD for optimized sensitivity estimation in mSENSE or recovering k-space directly in GRAPPA like methods. [1] R. Heckel and P. Hand, arXiv:1810.03982, Feb. 2019. [2] S. Arora, V. Roeloffs, and M. Lustig, in ISMRM 2020. [3] N. Weiskopf et al., Front Neurosci, vol. 7, p. 95, Jun. 2013.