Lung CT harmonization of paired reconstruction kernel images using generative adversarial networks

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Healthcare

Background:

• CT image reconstruction kernels significantly impact image texture and quantitative measurements.
• Inconsistent kernel choices across different scanners or protocols can introduce measurement variability unrelated to actual anatomy.

Purpose of the Study:

• To investigate and validate a deep learning approach for kernel harmonization in multi-vendor low-dose CT lung cancer screening data.
• To assess the impact of kernel conversion on quantitative CT-based assessments and radiomic feature reproducibility.

Main Methods:

• Utilized the National Lung Screening Trial dataset to identify paired CT scans reconstructed with soft and hard kernels.
• Employed the pix2pix deep learning architecture for kernel conversion, training 10 models on 100 pairs each.
• Evaluated conversion efficacy using image similarity metrics (RMSE, PSNR, SSIM) and assessed impact on emphysema, body composition, and radiomic features.

Main Results:

• Deep learning models effectively converted CT images between different kernel types, significantly improving image similarity metrics (RMSE, PSNR, SSIM).
• Kernel harmonization led to increased agreement in quantitative measurements of percent emphysema, skeletal muscle area, and subcutaneous adipose tissue (SAT) area.
• Radiomic features demonstrated improved reproducibility after kernel conversion compared to ground truth features.

Conclusions:

• Deep learning-based kernel conversion is a valid method for harmonizing CT images across different reconstruction kernels.
• This technique effectively reduces measurement variation in key quantitative metrics relevant to lung cancer screening and body composition analysis.