Interpolation method of CT image using features inferred by self-supervised learning

Abstract

Volume image is 3D data that is generated by stacking 2D slice images, and it has an advantage on 3D structure visualization and quantitative analysis. CT image is one of the volume images, and sometimes the image is acquired in low quality with long distance between slices to reduce noise and X-ray dose. Image interpolation is necessary for the low-quality image because it can be visualized discontinuously or cause an error in data analysis during image reconstruction to 3D. In this paper, we propose an interpolation method that infers and uses the information which is necessary for CT image interpolation from the high-resolution slice images through self-supervised learning. To do this, downscaled slice images are given as the input of the network, and the network recovers the original slice images from the downscaled images. The result of our method outperformed nearest-neighbor and trilinear interpolation and was similar to the supervised model with the same network. Among the networks, network using trilinear interpolation as the downscaling method and L2 loss with gradient loss produces the smoothest image. We hope this research contributes to the interpolation of volume data.