Purpose: To build a realistic digital phantom by synthesizing detailed anatomical structures and textures into the existing XCAT phantoms using cycle-GAN. As an initial feasibility study, we will focus on synthesizing CT textures for tumor and lung volumes.
Methods: This study has two main components: 1) Cycle-GAN model training: The model is built by simultaneously training two adversarial generator-discriminator pairs. The generators utilize ResNet structure and the discriminators utilize patchGAN structure. CT of 78 patients with stageâ… non-small cell lung cancer (NSCLC) were used. Tumor volume was contoured by physicians, and lung volumes were extracted using thresholding with vessels and bronchus excluded. CT images of tumor and lung and their corresponding binary mask images were used as input to train the model to synthesize textures in homogeneous mask images. 2) Generating XMAN phantom from XCAT: 10 XCAT lungs and 4 XCAT tumors of diameter(mm) [10, 20, 30, 40] were used for testing the Cycle-GAN. Homogeneous XCAT lung and tumor images were fed into the trained model to synthesize lungs and tumors with realistic textures. For evaluation, the mean, standard deviation, and histogram of the synthetic images were compared with those of real patient CTs.
Results: Qualitatively, textures generated in the synthetic lungs and tumors were comparable to real CT images. Quantitatively, the mean HU and standard deviation for XMAN lungs/tumors were 189.30Â±75.06/929.27Â±132.94, which were close to the values in real CTs (201.90Â±107.04/911.43Â±313.53). The histogram of synthetic images showed similar trend (probability distribution, peak, skewness) with that of CTs.
Conclusion: The preliminary study demonstrated feasibility of using Cycle-GAN to generate realistic CT textures in the XMAN phantom. Future studies will simulate MRI and PET image textures to introduce multi-contrast in the phantom. XMAN phantom can become valuable tools for technique evaluations or virtual trials in imaging and radiotherapy.