Anatomy-Guided Multi-Path CycleGAN for Lung CT Kernel Harmonization

Posted by Aravind Krishnan on Friday, December 5, 2025 in Computed Tomography, Deep Learning, Generative Adversarial Networks, Harmonization, Image Processing, Lung Cancer, Lung Screening CT.

Aravind Krishnan, Thomas Li, Lucas Walker Remedios, Kaiwen Xu, Lianrui Zuo, Kim L Sandler, Fabien Maldonado, Bennett Allan Landman, MIDL 2025, Link to paper: https://openreview.net/pdf?id=w3p7GddsQ8

 

Abstract 

Accurate quantitative measurement in lung computed tomography (CT) imaging often re-

lies on consistent kernel reconstruction across scanners and manufacturers. Harmonization

can reduce measurement variability caused by heterogeneous reconstruction kernels; how-

ever, harmonization across different manufacturers and scanners remains challenging due

to significant differences in reconstruction protocol and positional alignment of subjects,

often resulting in anatomical hallucinations. To address this, we propose a multi-path cy-

cleGAN framework that incorporates multi-region anatomical labels and a tissue statistic

loss as anatomical regularization to preserve structural integrity during harmonization. We

trained our model on 100 scans each of four representative reconstruction kernels from the

National Lung Screening Trial (NLST) dataset and evaluated it on 240 withheld scans.

Experimental results demonstrate superior performance of our method in both within-

manufacturer harmonization and cross-manufacture harmonization: Harmonizing hard-to

soft-kernel images within a single manufacturer significantly reduces emphysema measure-

ment discrepancies (p < 0.05). Across manufacturers, harmonizing all kernels to a reference

soft kernel yields consistent emphysema quantification (p > 0.05) and preserves anatomical

structures, as demonstrated by improved Dice similarity coefficient in skeletal muscle and

subcutaneous adipose tissue between harmonized and unharmonized images. These find-

ings demonstrate that segmentation-driven anatomical regularization effectively addresses

cross-manufacturer discrepancies, ensuring robust quantitative imaging. We release our

code and model at https://github.com/MASILab/AnatomyconstrainedMultipathGAN.

Keywords: cycleGAN, harmonization, CT, emphysema, synthesis

 

Figure: For any given pair of reconstruction kernels, the generator is a ResNet, formed

by a source encoder and target decoder in the forward path and a target encoder

and source decoder in the backward path. Each generator produces a synthetic

image with the style of either the source or target kernel. A PatchGAN is used

as a discriminator for the corresponding domain to distinguish between real and

synthetic images. The mean of all unique labels are computed using the multilabel

masks for the real and synthetic image and are penalized such that the anatomy

remains preserved in the harmonized image.