Reconstruction of Thin-Slice Medical Images Using Generative Adversarial Network

Abstract

Slice thickness is a very important parameter for medical imaging such as magnetic resonance (MR) imaging or computed tomography (CT). Thinner slice imaging obviously provides higher spatial resolution and more diagnostic information, however also involves higher imaging cost both in time and expense. For the sake of efficiency, a relatively thick slice interval is usually used in the daily routine medical imaging. A novel generative adversarial network was proposed in this paper to reconstruct medical images with thinner slice thickness from regular thick slice images. A fully convolutional network with three-dimensional convolutional kernels and residual blocks was firstly applied to generate the slices between the imaging intervals. A novel perceptual loss function was proposed to guarantee both the pixel similarity and the spatial coherence in 3D. Moreover, a discriminator network with a sustained adversarial loss was utilized to push the solution to be more realistic. 43 pairs of MR images were used to validate the performance of the proposed method. The presented method is able to recover preoperative t2flair MR images with slice thickness of 2 mm from routine t2flair MR images with thickness of 6 mm. The reconstruction results on two datasets show the superiority of the presented method over other competitive image reconstruction methods.