Author: Manju Liu, Weiwei Sang, Yanyan Shi, Zhenyu Yang, Fang-Fang Yin, Chulong Zhang, Lihua Zhang, Rihui Zhang 👨🔬
Affiliation: Jiahui International Hospital, Jiahui International Hospital, Radiation Oncology, Duke Kunshan University, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose: This study aims to transform cone-beam computed tomography (CBCT) images acquired from deep inspiration breath-hold (DIBH) breast cancer patients into high-fidelity synthetic CT (sCT) images. By mitigating CBCT artifacts and enhancing image quality, we seek to improve the accuracy of image-guided radiation therapy (IGRT).
Methods: We collected paired planning CT and CBCT data from 40 DIBH breast cancer patients, resulting in a total of 2880 image slices. Of these, 2001 slices were used for training and 879 for testing, ensuring that no slices from the same patient appeared in both sets. The unpaired CBCT-to-CT translation was achieved using an image-to-image network with shortest path regularization. We used PyANTs for registration and compared our method with CycleGAN and CUT. The performance was evaluated using structural similarity index measure (SSIM), mean absolute error (MAE), and peak signal-to-noise ratio (PSNR).
Results: Our approach achieved superior performance across all metrics. Specifically, for the sCT vs. registered planning CT comparison, our method yielded an average SSIM of 0.9652±0.0207, a PSNR of 30.9115±2.1251, and an MAE of 0.0053±0.0018. These results outperform the baseline CBCT (origin), CycleGAN, and CUT, demonstrating that the proposed method successfully enhances CBCT image quality and reduces artifacts.
Conclusion: The shortest path regularization-based unpaired image-to-image framework provides improved accuracy in generating sCT from CBCT images for DIBH breast cancer patients. The enhanced image quality and reduced artifacts have the potential to facilitate more precise IGRT, ultimately improving clinical workflows and treatment outcomes.