Generating synthesized computed tomography from CBCT/LDCT using a novel Generative-Transformer Adversarial-CNN

In the past few decades, many works have devoted to obtain low-dose CT (LDCT) or cone-beam CT (CBCT) to reduce the risk of X-rays in computed tomography (CT). Generating synthetic CT (sCT) images with low-dose examination (such as LDCT and CBCT) is an effective method. However, it is exceedingly challenging to generate high-quality sCT images from CBCT or LDCT images because of the poor image quality caused by artifacts, noise, complex multi-scale structure and surrounding tissue. In this article, we propose a novel Generative-Transformer Adversarial-CNN (GTAC) to achieve high-quality sCT images with low-dose examination. The GTAC learns a multi-scale mapping function to correct the HU value between low-dose CT and high-quality CT images. Specifically, the GTAC is composed of transformer-style UNet for sCT images generation and CNN-style Network for discriminator, the transformer-style UNet learns the global structures of anatomies in multiple scales with the well-designed. The Channel Transformer (CTrans) and Multi-Scale Feature Aggregation (MSFA) module capture the multi-scale structure information of anatomies and tissues, which makes the generated sCT images with complete anatomies and details. The Cascade Multi-Scale Convolution (CMSC) module trains the generator with low noise and artifacts by aggregating multi-scale information. The proposed GTAC is evaluated on two challenging dataset: private data from Xiangya Hospital of Central South University, public data from ‘2016 NIH AAPM-Mayo Clinic Low-dose CT Challenge’. The experimental results show the advantages of the proposed GTAC on sCT images generation with higher peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and lower mean absolute error (MAE), root mean squared error (RMSE).