CycleGAN for Kernel-to-Kernel CT harmonization and low-dose CT denoising via transfer learning

Low-dose CT (LDCT) reduces radiation exposure but increases image noise, while reconstruction kernel variability introduces texture inconsistencies. We propose a CycleGAN-based framework in which kernel harmonization is used as a pre-training task to provide a robust initialization for subsequent fine-tuning on LDCT denoising under unpaired and limited-data conditions.

A CycleGAN model was first trained to translate sharp into soft kernel reconstructions, learning transferable high-frequency texture representations. The pretrained model was then fine-tuned for LDCT denoising on two independent datasets (non-contrast chest and contrast-enhanced abdomen) using unpaired LDCT-NDCT data. Performance was evaluated on 40 paired test volumes per dataset using similarity metrics and high-frequency Noise Power Spectrum (NPS) correlation. Results were stratified by Body Area (BA) and compared with a conventional denoising approach (BM3D).

In non-contrast LDCT denoising, harmonization-based initialization improved structural similarity and maintained higher NPS correlation with NDCT, particularly for larger BA values associated with severe noise. In contrast-enhanced scans, baseline LDCT quality was already high, limiting gains in conventional metrics; however, harmonization-initialized CycleGAN achieved superior high-frequency NPS alignment.

Kernel harmonization serves as an effective pre-training task, providing robust initialization for subsequent fine-tuning in unpaired LDCT denoising and improving spectral fidelity under limited-data conditions.