Improved GAN: Using a transformer module generator approach for material decomposition. (October 2022)

• Record Type: Journal Article
• Authors: Wang, Guoshuai; Liu, Zhou; Huang, Zhengyong; Zhang, Na; Luo, Honghong; Liu, Lijian; Shen, Hao; Che, Canwen; Niu, Tianye; Liang, Dong; Luo, Dehong; Hu, Zhanli
• Is Part Of: Computers in biology and medicine. Volume 149(2022)
• Publication Date: 2022-10
• Subjects: Deep learning -- Generative adversarial network -- Transformer module -- Material decomposition -- Dual-energy CT; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2022.105952
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Dual-energy computed tomography (CT) can be used for material decomposition, allowing for the precise quantitative mapping of body substances; this has a wide range of clinical applications, including disease diagnosis, treatment response evaluation and prognosis prediction. However, dual-energy CT has not yet become the mainstream technique in most clinical settings due to its limited accessibility. To fully take advantage of material quantification, researchers have attempted to use deep learning to generate material decomposition maps from conventional single-energy CT images, mainly by synthesizing another single-energy CT image from a conventional single-energy CT image to form a dual-energy CT image first and then generate material decomposition maps. This is not a straightforward process, and it potentially introduces many inaccuracies after multiple steps. In this work, we proposed a generative adversarial network (GAN) framework as the base and improved its generator; this approach combines convolutional neural networks (CNNs) and a transformer module to directly generate material decomposition maps from conventional single-energy CT images. Our model pays attention to both local and global information. Then, we compared our method with 6 competitive deep learning methods on water (calcium) and calcium (water) substrate density image datasets. The average PSNR, SSIM, MAE, and RMSE of the generated and ground truth of the water (calcium) substrate density images were 32.7207, 0.9685, 0.0323, and 0.0555, respectively. Furthermore, the average PSNR, SSIM, MAE, and RMSE of the generated and ground truth of the calcium (water) substrate density images were 30.2823, 0.9449, 0.0652, and 0.0715, respectively. Our model achieved better performance and stronger stability than competing approaches. Highlights: We attempted to map conventional clinical single-energy CT images to material decomposition maps. We proposed an improved GAN model that combines a CNN and a Transformer as a generator. We first introduced the architecture of the Transformer in the material decomposition. Our model obtained superior performance to that of comparison methods in terms of image quality.