Residual CNN-based Image Super-Resolution for CT Slice Thickness Reduction using Paired CT Scans : Preliminary Validation Study

Woong Bae, Seungho Lee, Gwangbeen Park, Hyunho Park, Kyu-Hwan Jung

Abstract:

We propose a 2.5D image super resolution (SR) network based on fully residual
convolutional neural networks(CNN) which reduce the effective slice thickness of
CT scans. We demonstrate that the proposed network quantitatively outperforms
the 2017 NTIRE winning method, when trained and tested with 100 pairs of chest
CT scans acquired with different slice thickness (1 mm, 3 mm, 5 mm). Based on
the knowledge of CT reconstruction algorithms, we also demonstrate the necessity
of using real pairs of CT scans with different slice thickness rather than using
simulated low-resolution data which are widely used in image super resolution
studies. Furthermore, when the proposed SR method was applied, for CT images
that are 3mm and 5mm slice thickness, we confirmed dramatic performance
improvement in the CNN based lung nodule detection network.

Author Affiliation: VUNO Inc.

Keywords: Super Resolution, Fully Residual Convolutional neural networks, Computed Tomography (CT), Slice thickness