Scholarlink

<p>Cone-beam CT (CBCT) is widely used for patient positioning in image-guided radiotherapy (IGRT), but full-fan CBCT, which reduces patient dose and acquisition time, suffers from a smaller field of view (FOV) and degraded image quality. To address this limitation, we developed a generative adversarial network (GAN)-based image translation framework, specifically a Cycle-consistent GAN (CycleGAN), to enhance small-FOV CBCT images. The method incorporates dedicated preprocessing and multiple structure-preserving loss functions to mitigate boundary effects caused by truncation of patient anatomy. The model was trained and evaluated on CBCT and planning CT (PlanCT) images from prostate cancer patients. Experimental results demonstrated that the proposed approach substantially improved soft-tissue contrast, reduced noise, and preserved anatomical structures, as confirmed by Hounsfield unit statistics, histogram analysis, and noised SelfSSIM metrics. Importantly, the generated images exhibited image characteristics comparable to PlanCT, supporting accurate patient positioning. These findings indicate that the proposed GAN-based framework effectively enhances the image quality of small-FOV CBCT while maintaining anatomical fidelity, offering practical potential for clinical IGRT workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Image quality enhancement of small field-of-view cone-beam CT for image-guided radiotherapy using a structure-preserving cycle-consistent generative adversarial network

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Idzuru Yoshinaga,
  • Sho Ozaki,
  • Hideki Obara,
  • Masahiko Aoki

摘要

Cone-beam CT (CBCT) is widely used for patient positioning in image-guided radiotherapy (IGRT), but full-fan CBCT, which reduces patient dose and acquisition time, suffers from a smaller field of view (FOV) and degraded image quality. To address this limitation, we developed a generative adversarial network (GAN)-based image translation framework, specifically a Cycle-consistent GAN (CycleGAN), to enhance small-FOV CBCT images. The method incorporates dedicated preprocessing and multiple structure-preserving loss functions to mitigate boundary effects caused by truncation of patient anatomy. The model was trained and evaluated on CBCT and planning CT (PlanCT) images from prostate cancer patients. Experimental results demonstrated that the proposed approach substantially improved soft-tissue contrast, reduced noise, and preserved anatomical structures, as confirmed by Hounsfield unit statistics, histogram analysis, and noised SelfSSIM metrics. Importantly, the generated images exhibited image characteristics comparable to PlanCT, supporting accurate patient positioning. These findings indicate that the proposed GAN-based framework effectively enhances the image quality of small-FOV CBCT while maintaining anatomical fidelity, offering practical potential for clinical IGRT workflows.