Skip to main content

Inter-vendor harmonization of CT reconstruction kernels using unpaired image translation

Posted by on Friday, December 1, 2023 in Computed Tomography, Deep Learning, Generative Adversarial Networks, Harmonization, Image Processing, Lung Screening CT.

Aravind R. Krishnan, Kaiwen Xu, Thomas Li, Chenyu Gao, Lucas W. Remedios, Praitayini Kanakaraj, Ho Hin Lee, Shunxing Bao, Kim L. Sandler, Fabien Maldonado, Ivana Išgum, Bennett A. Landman. “Inter-vendor harmonization of CT reconstruction kernels using unpaired image translation” https://arxiv.org/abs/2309.12953

Abstract

The reconstruction kernel in computed tomography (CT) generation determines the texture of the image. Consistency in reconstruction kernels is important as the underlying CT texture can impact measurements during quantitative image analysis. Harmonization (i.e., kernel conversion) minimizes differences in measurements due to inconsistent reconstruction kernels. Existing methods investigate harmonization of CT scans in single or multiple manufacturers. However, these methods require paired scans of hard and soft reconstruction kernels that are spatially and anatomically aligned. Additionally, a large number of models need to be trained across different kernel pairs within manufacturers. In this study, we adopt an unpaired image translation approach to investigate harmonization between and across reconstruction kernels from different manufacturers by constructing a multipath cycle generative adversarial network (GAN). We use hard and soft reconstruction kernels from the Siemens and GE vendors from the National Lung Screening trial dataset. We use 50 scans from each reconstruction kernel and train a multipath cycle GAN. To evaluate the effect of harmonization on the reconstruction kernels, we harmonize 50 scans each from Siemens hard kernel, GE soft kernel and GE hard kernel to a reference Siemens soft kernel (B30f) and evaluate percent emphysema. We fit a linear model by considering the age, smoking status, sex and vendor and perform an analysis of variance (ANOVA) on the emphysema scores. Our approach minimizes differences in emphysema measurement and highlights the impact of age, sex, smoking status and vendor on emphysema quantification.

Differences in reconstruction kernels can be minimized by harmonizing to a reference standard. Harmonizing between paired kernels (left) has been explored due to the presence of one-to-one pixel correspondence between scans. However, unpaired kernels (right) create additional difficulties due to the difference in the anatomical alignment of scans obtained for different subjects from different vendors.
Fig 1. Differences in reconstruction kernels can be minimized by harmonizing to a reference standard. Harmonizing between paired kernels (left) has been explored due to the presence of one-to-one pixel correspondence between scans. However, unpaired kernels (right) create additional difficulties due to the difference in the anatomical alignment of scans obtained for different subjects from different vendors.

 

Kernel harmonization across four different reconstruction kernels can be performed using multiple cycle GANs operating across multiple paths. For a given source domain, a latent space is obtained from a source encoder that can be decoded by the corresponding target decoders. This approach enables harmonization between kernels from the same vendor and across kernels from different vendors using a high dimensional shared latent space (denoted as “L”).
Fig 2. Kernel harmonization across four different reconstruction kernels can be performed using multiple cycle GANs operating across multiple paths. For a given source domain, a latent space is obtained from a source encoder that can be decoded by the corresponding target decoders. This approach enables harmonization between kernels from the same vendor and across kernels from different vendors using a high dimensional shared latent space (denoted as “L”).

 

A cycle GAN consists of a forward and backward path. In the forward path, the source encoder and target decoder combine together to form a U-Net that generates a synthetic image with the style of the target domain. The synthetic image and the real target domain image are passed as inputs to a discriminator DB which distinguishes whether the generated image is real or fake. In the backward path, a synthetic image with the style of the source domain is generated which is fed to discriminator DA along with the source domain image.
Fig 3. A cycle GAN consists of a forward and backward path. In the forward path, the source encoder and target decoder combine together to form a U-Net that generates a synthetic image with the style of the target domain. The synthetic image and the real target domain image are passed as inputs to a discriminator DB which distinguishes whether the generated image is real or fake. In the backward path, a synthetic image with the style of the source domain is generated which is fed to discriminator DA along with the source domain image.

 

The noise in reconstruction kernels creates differences in the texture of underlying anatomical structures. The B50f and GE BONE hard kernels are noisy while the B30f and GE STD kernels are less noisy. Although the B30f and GE STD are soft kernels, their noise levels are different as these kernels belong to different vendors. Standardizing the GE soft kernel, GE BONE kernel and the B50f kernel to the reference B30f kernel (row 2) ensures consistent texture across all kernels for quantitative image analysis.
Fig 4. The noise in reconstruction kernels creates differences in the texture of underlying anatomical structures. The B50f and GE BONE hard kernels are noisy while the B30f and GE STD kernels are less noisy. Although the B30f and GE STD are soft kernels, their noise levels are different as these kernels belong to different vendors. Standardizing the GE soft kernel, GE BONE kernel and the B50f kernel to the reference B30f kernel (row 2) ensures consistent texture across all kernels for quantitative image analysis.

 

Percentage emphysema scores are affected by the reconstruction kernel in a given vendor, resulting in differences in measurements. Hard kernels overestimate emphysema quantification. Harmonizing kernels from different vendors to a reference soft kernel minimizes measurement errors, leading to a consensus among vendors for emphysema measurement.
Fig 5. Percentage emphysema scores are affected by the reconstruction kernel in a given vendor, resulting in differences in measurements. Hard kernels overestimate emphysema quantification. Harmonizing kernels from different vendors to a reference soft kernel minimizes measurement errors, leading to a consensus among vendors for emphysema measurement.

 

Emphysema quantification of GE BONE after kernel harmonization to B30f remained challenging. Although harmonization correctly reduced emphysema variation (left) in a few subjects, emphysema in the majority of subjects was over estimated (right).
Fig 6. Emphysema quantification of GE BONE after kernel harmonization to B30f remained challenging. Although harmonization correctly reduced emphysema variation (left) in a few subjects, emphysema in the majority of subjects was over estimated (right).