Accuracy enhancement of CT kernel conversions using convolutional neural net for super-resolution with Squeeze-and-Excitation blocks and progressive learning among smooth and sharp kernels

Ilsang Woo, Beomhee Park, Joon Beom Seo, Sang Min Lee, Namkug Kim

Abstract: Reconstructing smooth to sharp images with several CT kernels have been widely used in radiology, because images reconstructed with standard and high frequency kernels are more suitable for radiologic reading, but smoother kernels of higher SNR for quantitative software. However, it is very difficult to sustain several datasets with different kernels, due to limited storage and maintenance issues. We proposed accuracy enhancement of CT kernel conversion method using convolutional neural net for super resolution with SE block and progressive learning among smooth and sharp kernels. Our CT kernel conversion method showed significant enhancement and could be applicable to actual clinical environment for radiologic reading and quantitative SW without duplicated datasets.

Keywords: Super-Resolution, Squeeze-and-Excitation block, Progressive Learning

Author Affiliation: Asan Medical Center, University of Ulsan College of Medicine