3437 - Synthesis of Lung CT to Proton CT Based on CycleGAN

C. Sheng¹, Y. Ding¹, Y. Qi², M. Hu³, J. Zhang⁴, X. Cui⁵, Y. Zhang⁶, and W. Huo¹; ¹Key Laboratory of Electromagnetic Wave Information Technology and Metrology of Zhejiang Province, College of Information Engineering, China Jiliang University, Hangzhou, Zhejiang, China, ²Division of lonizing Radiation Metrology, National Institute of Metrology, Beijing, Beijing, China, ³Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, ⁴Departments of Radiation Oncology, Zibo Wanjie Cancer Hospital, Zibo, Shandong, China, ⁵Institute of Health and Medical Technology, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, Anhui, China, ⁶Department of Oncology, Xiangya Hospital, Central South University, Changsha, Hunan, China

Purpose/Objective(s): Proton radiography is an imaging method using proton beam instead of X ray to get patient image. Water equivalent path length (WEPL) information can be directly got in proton radiograph, which is critical to proton therapy. Deep learning has been widely applied in medical image field due to its powerful feature extraction capability. However, proton radiography is a frontier medical image technology, the sample size of proton image dataset is too small to use to train deep learning network. Computed tomography (CT) is the most common examination method in clinic. In this paper, we use a improved cycle generative adversarial network (CycleGAN) to convert unpaired and uncalibrated lung CTs into proton CTs to expand the sample size of lung proton CTs dataset to make it more suitable for deep learning network training. Materials/

Methods: We selected 3D CT data from 53 lung cancer patients and 4D proton CT data including 10 different phases from 1 lung cancer patient for preprocessing. Firstly, all the data were sliced into 2D data and cropped to appropriate sizes, and then high-quality images were filtered. The training dataset included unpaired and uncalibrated 832 CT slice images and 1185 proton CT slice images. We constructed a CycleGAN to demonstrate the possibility of using it to expand proton CTs through CTs. Subsequently, we added a sobel operator convolutional layer to CycleGAN encoder to extract edge and detail features. A structural similarity (SSIM) loss was added in loss function to make generated proton CT structure more similar to that of CT. In addition, convolutional layers with 1 kernel size were used to replace upsampling and downsampling layer, to make generated proton CTs style feature more closer to that of real proton CTs. Finally, 15 CT slice images were randomly selected from rest patient CT as testing dataset to test the performance of above CycleGANs and generated proton CTs were evaluated.
Results: The Frechet Inception Distance (FID) score and Turing like test score of proton CT generated by original CycleGAN was 255.50 and 2.63, respectively. The proton CTs generated by the CycleGAN which was added a sobel operator convolutional layer in generator and SSIM loss in the loss function had a 220.00 score for FID and a 2.90 score for Turing like tests. After updating upsampling and downsampling layer, the FID score of generated proton CT decreased to 188.22, and its Turing like test score increased to 4.10. The Turing like test score was significantly improved after used improved CycleGAN. Even if the FID score didn’t decrease significantly, this improvement also made generated proton CT style feature, especially grayscale values more closer to real proton CTs.
Conclusion: The improved CycleGAN had feasibility and effectiveness in expanding proton CTs. It can generate proton CTs that are similar to real samples in terms of structure, detail information, style features, etc., laying the foundation for the application of artificial intelligence in proton CT.