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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Quantitative Image Analysis

Background:

• Reconstruction kernels in computed tomography (CT) significantly influence image texture, impacting quantitative analysis.
• Inconsistent CT textures across different kernels and manufacturers lead to measurement discrepancies.
• Existing harmonization methods often require paired, aligned scans and extensive model training.

Purpose of the Study:

• To develop and evaluate an unpaired image translation approach for harmonizing CT reconstruction kernels.
• To investigate harmonization between and across different manufacturers' kernels using a multipath cycle generative adversarial network (GAN).
• To assess the impact of harmonization on quantitative measurements, specifically percent emphysema.

Main Methods:

• Utilized a multipath cycle generative adversarial network (GAN) for unpaired image translation.
• Trained the GAN on hard and soft reconstruction kernels from Siemens and GE vendors using the National Lung Screening Trial dataset.
• Harmonized scans to a reference Siemens soft kernel (B30f) and evaluated percent emphysema using linear modeling and ANOVA.

Main Results:

• The multipath cycle GAN successfully harmonized CT scans across different reconstruction kernels and manufacturers.
• Harmonization minimized differences in percent emphysema measurements.
• Analysis revealed significant impacts of age, sex, smoking status, and vendor on emphysema quantification.

Conclusions:

• The proposed unpaired image translation approach effectively harmonizes CT reconstruction kernels, reducing measurement variability.
• This method facilitates more consistent quantitative image analysis across diverse CT datasets.
• The study underscores the importance of accounting for patient demographics and scanner vendor in emphysema quantification.