Synthetic CT Imaging Using Limited Patch-Based Cyclegan for Adaptive Radiotherapy
Abstract
Purpose
Cycle generative adversarial network (GAN) has a difficulty in maintaining the structural integrity in a synthesized image for complex organs and tissues. In this study, we proposed a CycleGAN including multi-scale blocks and attention gates for improving the accuracy of synthetic CT images in adaptive radiotherapy. And, the location-limited patches were used for precisely calculating a loss during network training.
Methods
The generator of the proposed CycleGAN had a U-Net structure, and convolution layers were replaced with 9 multi-scale blocks for securing the diversity in extracted features and enhancing network robustness. 4 attention gates were added between the former and latter blocks for precisely delivering the crucial information for a synthetic task. 5 convolution layers were used for a discrimination procedure. 8 patches were extracted from both of ground truth (GT) and synthetic CT images, and the extraction locations were limited to include an object and reject a background. The difference between the extracted patches were calculated by passing through the encoder part of the generator, two convolution layers and up-sampling layer. The quantitative accuracy of the output CT images was compared with the Pix2Pix, conventional CycleGAN and CUT models in terms of PSNR, SSIM and NRMSE.
Results
Compared to the other models, the visual similarity between the GT and output CT images was highest for the proposed model. The PSNR and SSIM for the proposed model were 3.88-6.05 and 5.81-14.70% higher those for the other models. And, the proposed model reduced the NRMSE of the output CT image by 32.96-37.81% compared to the other models.
Conclusion
The proposed CycleGAN model can be potentially utilized for improving the quantitative accuracy of synthetic CT images. And the proposed model is expected to precisely provide the anatomical information of a patient for adaptive radiotherapy.