PPubMed8 Nov 2025

Multipath cycleGAN for harmonization of paired and unpaired low‐dose lung computed tomography reconstruction kernels KRISHNAN et al.

Aravind R. Krishnan, Thomas Z. Li, Lucas W. Remedios, Michael E. Kim, Chenyu Gao, Gaurav Rudravaram, Elyssa M. McMaster, Adam M. Saunders, Shunxing Bao, Kaiwen Xu, Lianrui Zuo, Kim L. Sandler, Fabien Maldonado, Yuankai Huo, Bennett A. Landman

Abstract

5## DISCUSSION AND CONCLUSION In this work, we investigate paired and unpaired CT kernel harmonization in a low-dose lung cancer screening cohort spanning multiple vendors. We develop a two-stage multipath cycleGAN model with a shared latent space, enabling harmonization across different combinations of paired and unpaired reconstruction kernels within a unified framework. The proposed model is evaluated against cycleGAN and switchable cycleGAN models for percent emphysema quantification and anatomical consistency in unpaired kernels. Additionally, we investigate the impact of kernel on percen

A figure from Multipath cycleGAN for harmonization of paired and unpaired low‐dose lung computed tomography reconstruction kernels KRISHNAN et al. — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

A multipath cycleGAN with a shared latent space harmonizes paired and unpaired CT reconstruction kernels, reducing emphysema quantification errors while preserving anatomy.

What’s new

i.Introduced a multipath cycleGAN architecture with a shared latent space for both paired and unpaired kernel harmonization.
ii.Demonstrated improved emphysema quantification across multiple vendors and reconstruction kernels.
iii.Maintained anatomical consistency (muscle, adipose tissue, lung vessels) after harmonization.
iv.Enabled joint training on paired and unpaired data using domain‑specific encoder‑decoder paths and multiple loss functions.

How they did it

The model was trained in two stages using paired kernels from the same manufacturer and unpaired kernels from different manufacturers, stitching domain‑specific encoder‑decoder networks through a shared latent space. Losses included adversarial, cycle‑consistency, identity L2, and a down‑sampled MSE to preserve radiopacity. Training used axial 512×512 slices with a binary mask to standardize the field‑of‑view.

What they found

For paired kernels, the multipath approach significantly reduced percent emphysema differences (p<0.05) and achieved the best performance on one kernel pair over standard cycleGAN.
On the (BONE, STANDARD) pair, multipath cycleGAN achieved a 91.52% reduction in median RMSE, the largest among methods for that pair.
For unpaired kernels harmonized to a reference soft kernel, the method mitigated differences in 3 of 6 kernels, comparable to switchable cycleGAN.
Anatomical consistency was maintained for muscle and adipose tissue across all unpaired kernels; lung‑vessel overlap was reasonable but lower than switchable cycleGAN.

Caveats

✕Performance on lung‑vessel segmentation was lower compared to the switchable cycleGAN.
✕Mitigation of kernel effects was comparable but not superior to switchable cycleGAN for several unpaired kernels.

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

Multipath cycleGAN for harmonization of paired and unpaired low‐dose lung computed tomography reconstruction kernels KRISHNAN et al.

The paper, in five minutes

Ask a question about this paper

Similar papers in all four archives

The paper, in five minutes

Similar papers in all four archives