Generative Adversarial Network-Based Image Conversion Among Different Computed Tomography Protocols and Vendors: Effects on Accuracy and Variability in Quantifying Regional Disease Patterns of Interstitial Lung Disease

The paper aimed to enhance the accuracy and consistency of quantifying interstitial lung disease (ILD) in computed tomography (CT) scans obtained from various manufacturers and scan settings. It employed a routable generative adversarial network (RouteGAN) to convert CT images from different acquisition conditions to a target style. Coreline Soft's deep learning-based software, Aview, was employed for automated quantification of ILD. The study involved CT images from 150 patients with ILD, obtained from different scanners, radiation doses, and kernel settings, categorized into groups. The converted CT images showed improved accuracy in quantifying total abnormalities, fibrosis score, honeycombing, and reticulation, as indicated by metrics such as dice similarity coefficient and pixel-wise overlap accuracy. Radiologists' evaluation also showed higher and less variable scores on converted CT images. The findings suggest that RouteGAN-based CT conversion can enhance the precision of ILD quantification across varied acquisition conditions, benefiting deep learning-based analysis like Aview by Coreline Soft.