Abstract: In this work we combine a mixed-scale dense convolutional network and a structure
preserving loss function, to increase the in-plane resolution of MRI images with
sub-millimeter resolution. Despite having 20 times fewer parameters than SRCNN,
this architecture can reconstruct high-resolution images in comparable quality
and learns to better preserve high-frequency details than networks trained with
L 2 -loss. Preliminary experiments show that the architecture in general is working
well and comparable to the well known SRCNN without any optimization of
hyper-parameters.
Keywords: Super-Resolution, Deep Learning, Convolutional Neural Networks, MRI
Author Affiliation: Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany
3 Replies
Loading