Synthesis of 3D Realistic High-resolution Lung Background Textures Using a Conditional Generative Adversarial Network (CGAN)

Abstract

Objectives: We develop machine-learning based methods to synthesize lung textures within computational phantoms for improved realism in simulating high-resolution patient CT imaging data to evaluate and improve imaging devices and techniques.Methods: We first optimized a previously developed technique designed using a Conditional Generative Adversarial Network (CGAN), Project 1. The optimized model was trained and validated using clinical CT data. Generated texture images were evaluated qualitatively and quantitatively comparing them to the original CT data as well as to results from the previous work. Using what we learned from Project 1, in Project 2, we trained and validated a new generator using high-resolution micro-CT data of the lungs. The new generator was evaluated in a similar fashion. Results: For Project 1, the model was unable to produce results better than the previous work; lung textures were found to be blurry and lacked detail. For Project 2, the trained generator was found capable of simulating variable 3D lung background textures similar to the micro-CT both qualitatively and quantitatively. Conclusion: The CGAN method developed in this work, based on micro-CT data, can greatly improve the realism of computational phantoms by adding high-resolution background textures to the lungs. Such anatomical detail is necessary to evaluate higher-resolution CT imaging methods such as photon-counting CT.