Abstract
Stereotactic body radiation therapy (SBRT) is a highly effective treatment for lung cancer; however, challenges arise from tumor motion induced by respiration. The CyberKnife system, incorporating both fiducial-based and fiducial-free tracking modalities, aims to mitigate these challenges, yet tumor recognition can be compromised by overlapping bone structures. This study introduces a novel bone suppression imaging technique for kilovolt X-ray imaging using generative adversarial networks (GANs) to enhance tumor tracking in SBRT by reducing interference from bony structures. Computed tomography (CT) images, both with and without bone structures, were generated using a four-dimensional extended cardiac-torso phantom (XCAT phantom) across 56 cases. X-ray projections were captured from left and right oblique 45° angles and divided into nine segments, producing 1120 images. These images were processed through six pre-trained GAN models-CycleGAN, DualGAN, CUT, FastCUT, DCLGAN, and SimDCL-yielding bone-suppressed images on the XCAT phantom (BSI(phantom)). The resulting images were evaluated against bone-shadow-free images using structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), and Frechet inception distance (FID). Additionally, bone-suppressed images (BSI(patient)) were derived from 1000 non-simulated patient images. BSI(phantom) images achieved SSIM and PSNR values of 0.96 ± 0.02 and 36.93 ± 3.93, respectively. SimDCL exhibited optimal performance with an FID score of 68.93, indicative of superior image generation quality. This GAN-based bone suppression imaging technique markedly improved image recognition and refined dynamic tumor tracking, enhancing the accuracy and efficacy of SBRT.