106P Generation of synthetic ground glass opacities (GGOs) using generative adversarial networks (GANs)

Abstract

The robustness of deep learning (DL)-based computer aided diagnosis (CAD) systems for the automated detection of pulmonary nodules in low-dose CT scans strongly relies on the availability of a large amount of curated and annotated data. Even when this holds true, the unbalance problem will exist. For example, GGOs will be harder to be recognized by the algorithm, because of their lower prevalence compared to part solid and solid nodules. DL algorithms such as GANs can be utilized to generate synthetic samples, thus increasing the original datasets and improve the CAD detection accuracy. However, the quality of the generated samples should be evaluated both qualitatively (visual inspection) as well as quantitively to prove a clinical relevance. We proposed and evaluated a novel structure for generating GGOs through Super resolution GANs. 216 CT images (340 GGO nodules) from The Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) database were used. An optimized super resolution GANs built on retrieving information from the whole image and the GGO region was developed. The generated samples were evaluated asking four trained radiologist/pulmonologist to perform a visual Turing test (VTT) and by comparing quantitative imaging features (radiomics) between real and synthetic nodules in four grades confidently fake, leaning fake, leaning real, and confidently real. The results of the VTT. More than one-third (37.3%) of the synthetic GGOs were classified as ‘Real’, of which 15.7% were classified as ‘Confidently Real. A total of 58/93 radiomic features had no significant difference (P > 0.05) between synthetic and real GGOs. The ROC curves constructed based on the results of radiologists and radiomics a similar classification performance between radiologists (AUC = 0.68) and radiomics (AUC = 0.66). In this study, we used GANs to generate synthetic GGOs in low-dose CT images. The above studies show that the generated nodules closely match real samples, both visually and quantitatively.