Sparse Coding Super-Resolution Scheme for Chest Computed Tomography

High-resolution chest computed tomography images now has a great importance in the diagnosis. However, this modality requires using a higher radiation dose and a longer scanning time compared to low-resolution computed tomography. In this study, we applied the sparse coding super-resolution method to reconstruct high-resolution images without increasing the radiation dose. We prepared an over-complete dictionary by mapping between low- and high-resolution patches and represented this as a sparse linear combination of each patch of the low-resolution input. These coefficients were used to reconstruct the high-resolution output. In our experiments, 89 computed tomography scans were analyzed. We up-sampled the images 2 or 4 times and compared the image quality of the sparse coding super-resolution scheme with those of the nearest neighbor and bilinear interpolations, which are the traditional interpolation schemes. The image quality was evaluated by measuring the peak signal-to-noise ratio and structure similarity. The differences in the peak signal-to-noise ratios and structure similarities between the sparse coding super-resolution method and the nearest neighbor or bilinear method were statistically significant. Visual assessment confirmed that the sparse coding super-resolution method generated high-resolution images, whereas the conventional interpolation methods generated over-smoothed images. Taken together, these results suggest that the sparse coding super-resolution approach is a robust method for up-sampling computed tomography images and that it yields images with markedly high resolution when magnifying chest computed tomography scans