Mamba Gan–Driven Synthetic CT Generation from CBCT for Online Adaptive Proton Therapy
Abstract
Purpose
CBCT is routinely acquired prior to proton therapy for patient setup. However, the limited image quality of CBCT compromises the dose calculation accuracy and limits its use for treatment plan adjustments. This study aims to develop a high-performance CBCT-to-CT synthesis framework based on the Mamba architecture, capable of generating high-quality synthetic CT (sCT) images, and establishing an online adaptive proton therapy workflow.
Methods
A GAN network based on the Mamba architecture was constructed to synthesize CBCT-to-CT images. The generator in the GAN network employed a U-Net with the Mamba architecture, while the discriminator used a five-stage down-sampling network structured based on Mamba. Network training incorporated adversarial, MAE, and histogram losses. The study included data from 52 patients undergoing proton therapy at our institution. Thirty-two whole-brain and head-and-neck cases were used for training, while 30 head-and-neck cases were used for testing. The evaluation indicators include mean absolute error (MAE), structural similarity index measurement (SSIM), and peak signal-to-noise ratio (PSNR).
Results
Our model significantly improves CBCT image quality while preserving the original anatomical structures. The quantitative values of MAE, PSNR, and SSIM were changed from 317.4±34.0 HU, 21.5±1.0 dB, and 0.918±0.016 to 77.2±13.5 HU, 39.4±0.4 dB, and 0.976±0.009, respectively. By integrating image synthesis, deformation registration, contour transfer, and dose accumulation, a workflow suitable for online proton plan evaluation has been established.
Conclusion
We developed a GAN network based on the Mamba architecture for CBCT-to-CT image synthesis. The image quality of CBCT has been significantly enhanced. The excellent image quality of the sCT demonstrates its potential to enable online adaptive proton therapy. After integrating image synthesis and registration, an online proton plan evaluation workflow has been established. Funding: This work was supported by the National Natural Science Foundation of China (No. 12375359) and CAMS Innovation Fund for Medical Sciences (CIFMS, 2024-I2M-C&T-B-076).