Multipath cycleGAN for harmonization of paired and unpaired low-dose lung computed tomography reconstruction kernels

Reconstruction kernels in computed tomography (CT) introduce variability in spatial resolution and noise distribution, creating systematic differences in quantitative imaging measurements, that icl, emphysema characterization in lung imaging. While harmonization across images reconstructed using different kernels from the same manufacturer is feasible, this can be challenging in multi-centre or longitudinal studies where images are acquired using kernels from different manufacturers. This variability results in heterogeneous quantitative measurements that are difficult to compare. Therefore, it is necessary to standardize all images to a common reference kernel.

We explore training a harmonization model using paired reconstruction kernels (obtained from the same manufacturer for a given subject) and unpaired reconstruction kernels (obtained across different manufacturers for different subjects) in a low dose lung cancer screening cohort, validating our approach through quantitative CT measurements. Our overall goal is to use both sets of data to construct a shared latent space while recognizing the ability of different types of data to contribute through different classes of loss functions.

We develop a multipath cycleGAN that enables multi-domain kernel harmonization through a shared latent space, domain specific encoder-decoder architectures, and discriminators trained in an unsupervised manner using a mixture of paired and unpaired data. We train our model using 100 scans each from seven representative kernels (Siemens B50f, Siemens B30f, GE BONE, GE STANDARD, GE LUNG, Philips C, and Philips D) from the National Lung Screening Trial (NLST) dataset, enabling harmonization across 42 different kernel combinations. Using 240 withheld scans from each kernel, we evaluate our approach on paired kernels using percent emphysema. For unpaired kernels, we harmonize all scans to the style of a reference soft kernel (Siemens B30f) and evaluate our model using percent emphysema, followed by a general linear model analysis that investigates the impact of age, sex, smoking status, and kernel on emphysema. Additionally, we harmonize all soft kernels to a reference hard kernel (Siemens B50f), quantifying percent emphysema. We assess anatomical consistency in unpaired kernels harmonized to a reference soft kernel by comparing segmentations of lung vessels, muscle, and subcutaneous adipose tissue derived from TotalSegmentator between the non-harmonized and harmonized images. We compare the performance of our model to the standard cycleGAN and a switchable cycleGAN model.

For paired kernels, the proposed multipath approach reduced differences in percent emphysema as seen from the Bland Altman analysis (p < 0.05), achieving the best performance on one kernel pair over the standard cycleGAN and on two pairs over the switchable cycleGAN.For unpaired kernels where all source kernels were harmonized to a reference soft kernel, our method mitigated differences in three of six kernels, comparable to the switchable cycleGAN, while the standard cycleGAN mitigated differences in four of six kernels. When harmonizing all soft kernels to a reference hard kernel, our approach outperformed the switchable cycleGAN and was comparable to the standard cycleGAN. Our proposed approach maintains anatomical consistency in muscle, adipose tissue across all unpaired kernels and showed reasonable overlap on three kernels for lung vessels when compared to the cycleGAN; however, performance was lower when compared to the switchable cycleGAN.

Paired and unpaired kernel harmonization with a shared latent space multipath cycleGAN mitigates errors in emphysema quantification and maintains anatomical consistency after harmonization.