Introduction Cone-beam computed tomography (CBCT) is instrumental in image-guided radiation therapy (IGRT), providing low radiation exposure while continuously monitoring anatomical structures for accurate dose estimation and treatment. Despite these advantages, CBCT inherently suffers from lower image quality and more frequent artifacts compared to computed tomography (CT), significantly undermining its effectiveness in IGRT. These drawbacks are especially pronounced in the pelvic region, where anatomical variability and dataset asymmetry challenge traditional image translation techniques like diffusion and CycleGAN networks. Methods To overcome these limitations, we propose CycleDiffSmoothGAN(CDSGAN), an innovative framework that enhances CBCT images by integrating CycleGAN with latent diffusion techniques and high-frequency detail preservation.This approach effectively blends features in the latent space, enabling smoother transitions between CBCT and synthetic CT (sCT) images. Results The implementation of CDSGAN has shown superior performance, significantly outperforming existing technologies across crucial imaging metrics such as MAE (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="234_2025_3634_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="87" /> </InlineMediaObject> <EquationSource Format="TEX">$18.53\pm 3.58$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>18.53</mn> <mo>±</mo> <mn>3.58</mn> </mrow> </math></EquationSource> </InlineEquation> Hu), PSNR (<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="234_2025_3634_Article_IEq2.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="87" /> </InlineMediaObject> <EquationSource Format="TEX">$26.90 \pm 1.53$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>26.90</mn> <mo>±</mo> <mn>1.53</mn> </mrow> </math></EquationSource> </InlineEquation> dB), SSIM (<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="234_2025_3634_Article_IEq3.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="80" /> </InlineMediaObject> <EquationSource Format="TEX">$0.90 \pm 0.03$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.90</mn> <mo>±</mo> <mn>0.03</mn> </mrow> </math></EquationSource> </InlineEquation>), and FID (<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="234_2025_3634_Article_IEq4.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="80" /> </InlineMediaObject> <EquationSource Format="TEX">$9.84\pm 1.21$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>9.84</mn> <mo>±</mo> <mn>1.21</mn> </mrow> </math></EquationSource> </InlineEquation>). Conclusion The research findings have substantiated the promising potential of CDSGAN in enhancing image quality for clinical applications.

CBCT-to-CT synthesis with a hybrid of CycleGAN and latent diffusion

期刊论文
发表日期：

2025年5月7日

查看全文

Feng Luo,
Chaoyu Ma,
Juntian Shi,
Kunyuan Xu

摘要

Introduction

Cone-beam computed tomography (CBCT) is instrumental in image-guided radiation therapy (IGRT), providing low radiation exposure while continuously monitoring anatomical structures for accurate dose estimation and treatment. Despite these advantages, CBCT inherently suffers from lower image quality and more frequent artifacts compared to computed tomography (CT), significantly undermining its effectiveness in IGRT. These drawbacks are especially pronounced in the pelvic region, where anatomical variability and dataset asymmetry challenge traditional image translation techniques like diffusion and CycleGAN networks.

Methods

To overcome these limitations, we propose CycleDiffSmoothGAN(CDSGAN), an innovative framework that enhances CBCT images by integrating CycleGAN with latent diffusion techniques and high-frequency detail preservation.This approach effectively blends features in the latent space, enabling smoother transitions between CBCT and synthetic CT (sCT) images.

Results

The implementation of CDSGAN has shown superior performance, significantly outperforming existing technologies across crucial imaging metrics such as MAE ( $18.53\pm 3.58$ $18.53 \pm 3.58$ Hu), PSNR ( $26.90 \pm 1.53$ $26.90 \pm 1.53$ dB), SSIM ( $0.90 \pm 0.03$ $0.90 \pm 0.03$ ), and FID ( $9.84\pm 1.21$ $9.84 \pm 1.21$ ).

Conclusion

The research findings have substantiated the promising potential of CDSGAN in enhancing image quality for clinical applications.