Megavoltage CT enhancement for cervical cancer tomotherapy using a generative adversarial network with deformable convolution and self‐attention
Abstract
Background
Megavoltage computed tomography (MVCT) is an essential imaging modality for verifying patient positioning in helical tomotherapy. However, its clinical application in daily anatomical monitoring and adaptive radiotherapy is hindered by inherent image artifacts and poor soft‐tissue contrast. This issue is particularly pronounced in pelvic radiotherapy, where the intra‐ and interfraction anatomical variations necessitate high‐quality image guidance to ensure precise dose delivery.
Purpose
We developed a deep‐learning‐based framework for MVCT enhancement to improve anatomical visualization and facilitate accurate adaptive treatment planning for cervical cancer.
Methods
This study analyzed a retrospective cohort of 170 patients with cervical cancer who underwent helical tomotherapy. The proposed deep‐learning‐based algorithm employed a generative adversarial network (GAN) that integrated deformable convolution and a self‐attention mechanism (SADC‐EGAN) to improve MVCT image quality. Comparative analyses were conducted against representative baseline methods, including U‐Net, Attention U‐Net, U‐Net++, Swin‐UNet, CycleGAN, and Pix2Pix. Model performance was assessed using quantitative metrics, including mean absolute error (MAE), peak signal‐to‐noise ratio (PSNR), structural similarity index measure (SSIM), and Fréchet inception distance (FID).
Results
The synthetic computed tomography (sCT) images generated by the proposed SADC‐EGAN method demonstrated superior Hounsfield unit (HU) accuracy and structural similarity compared to the original MVCT. Specifically, the MAE between the sCT and kilovoltage computed tomography (kVCT) was reduced to 36.32 ± 6.69 HU, compared with 56.72 ± 9.09 HU for MVCT. In terms of image quality, the sCT images exhibited notable enhancements over MVCT images, with higher PSNR (32.54 ± 2.31 vs. 29.40 ± 1.56 dB), improved SSIM (0.93 ± 0.01 vs. 0.89 ± 0.02), and substantially lower FID (66.68 ± 22.31 vs. 153.52 ± 28.77).
Conclusions
The proposed SADC‐EGAN framework, integrating deformable convolutions and self‐attention, effectively generated high‐quality kVCT‐like images from MVCT, improving both HU accuracy and image quality. This approach has clinical potential to enable online adaptive helical tomotherapy for cervical cancer.