To compare with fan-beam computed tomography (FBCT), the image quality of Cone-beam computed tomography (CBCT) is indistinct due to X-ray scatter contamination and truncated projections, which misses numerous effective information. We proposed a novel cycle-consistent adversarial network model method combined Cycle-GAN and Deblur-GAN, so-called the Cycle-Deblur GAN, to improve the image quality of CBCT for truncated chest CT images. The generated CBCT images from our proposed method demonstrated closer Hounsfield unit (HU) value to FBCT in lung, breast, mediastinum, and sternum compared to Cycle-GAN and RED-CNN. The proposed method evaluated by the mean absolute error (MAE), peak-to-signal noise ratio (PSNR), and the structural similarity index measure (SSIM) demonstrated better results. The Cycle-Deblur GAN improved image quality and preserved structural details for truncated chest CBCT images.

Our proposed Cycle-Deblur GAN consists of Cycle-GAN and Deblur-GAN with increasing the shortcut number and inception block to preserve the detail structure. For the activation layer, since the performance of Swish was better than LeakyReLU and ReLU, we adopted it as our activation function for Cycle-Deblur GAN. When using RED-CNN model to train CBCT and FBCT, the alignment problem existed and the results of RED-CNN model showed the blurring results. The image quality of generated CBCT images was improved and evaluated by HU value, PSNR, and SSIM. The artifact of CBCT was well removed by using this method. The proposed method can enhance the structural details in the lung, soft tissue, and bony structure so that the visualization enhanced and was more similar to FBCT.

The FBCT acquired from CT simulation were used as the ground truth images. The image quality of the CBCT was improved by the Cycle-Deblur GAN modeling. The dose distribution has become more and more compliant and steeper for these years, and the accuracy of the treatment position has become more important. Image registration between CBCT and FBCT before treatment was the most common image-guided radiotherapy technique. Since the scattering and artifacts removed after Cycle-Deblur GAN modeling, the radiotherapy could be more accurate. The CT number of each tissue was also closer to the FBCT through Cycle-Deblur GAN. For recent years, the radiomics research become more and more popular. The high image quality could increase the probability for radiomics research using CBCT.