Room: Exhibit Hall | Forum 9
Purpose: To use unpaired image-to-image translation capabilities of Cycle Consistent Adversarial Networks (CycleGAN) to generate synthetic CT images from CBCT images in treatments areas with higher inter-fraction variability, such as the prostate anatomy.
Methods: 82 CBCT, and CT images of patients who underwent prostate radiotherapy were anonymized, resampled and cropped to create the training data sets for the generative network and the discriminative network respectively. Both training sets had a resolution 0.9 x 0.9 x 1.99 mmÂ³ and dimensions of 256 x 256 x 88. A cycle-consistent adversarial network was implemented using TensorFlow (https://www.tensorflow.org) to achieve unpaired image-to-image translation between the two datasets. The loss function was modified to reduce ring artifacts and to further preserve HU values from the CT dataset.
Results: We compared the synthetically generated CT (GAN-CT) with a deformed CT (DCT) generated using the deformable registration toolkit from Eclipse (Varian Medical System). Specifically, we focused on three scenarios: normal image, image with seeds or fiducials and images with air. Mean Absolute Error (MAE), root mean squared error (RMSE), signal-to-noise ratio (SNR) and Peak SNR (PSNR) were used to assess image quality. Overall, GAN-CT showed higher accuracy than CBCT when compared with DCT. Line profiles displayed higher contrast for GAN-CT. Finally, linear regression of the GAN-CT signal intensity versus DCT showed improved signal intensity correction of the images.
Conclusion: Deformable registration is the most common method to correct Hounsfield unit values in CBCT images. However, this method is inaccurate when air pockets are found in the CBCT images or around metal artifacts. CycleGAN-based methods could possibly overcome this issue by preserving the structure of the CBCT and reducing the ring artifacts, while adopting the HU values of the planning CT