Low-dose CT Enhancement Network with a Perceptual Loss Function in the Spatial Frequency and Image Domains

TL;DR

This paper introduces a dual-domain U-net cascade for low-dose CT image enhancement, leveraging both spatial frequency and image domains to improve image quality without proprietary data, outperforming single-domain methods.

Contribution

The novel dual-domain W-net architecture combines spatial frequency and image domain processing, demonstrating superior performance over single-domain networks in LDCT enhancement.

Findings

Dual-domain W-net outperforms single-domain U-net cascades.

Spatial frequency domain learning enhances image quality.

Deep learning methods surpass traditional FBP in LDCT denoising.

Abstract

We propose a dual-domain cascade of U-nets (i.e. a "W-net") operating in both the spatial frequency and image domains to enhance low-dose CT (LDCT) images without the need for proprietary x-ray projection data. The central slice theorem motivated the use of the spatial frequency domain in place of the raw sinogram. Data were obtained from the AAPM Low-dose Grand Challenge. A combination of Fourier space (F) and/or image domain (I) U-nets and W-nets were trained with a multi-scale structural similarity and mean absolute error loss function to denoise filtered back projected (FBP) LDCT images while maintaining perceptual features important for diagnostic accuracy. Deep learning enhancements were superior to FBP LDCT images in quantitative and qualitative performance with the dual-domain W-nets outperforming single-domain U-net cascades. Our results suggest that spatial frequency learning in conjunction with image-domain processing can produce superior LDCT enhancement than image-domain-only networks.