4DCBCT-4DsCT Generation Via Diffusion Model In Lung Cancer Adaptive Radiotherapy
Abstract
Purpose
In lung cancer adaptive radiation therapy (ART), respiratory motion and interfractional anatomical variations—including changes in tumor size and normal tissue anatomy—can affect the efficacy of radiation treatment. Acquiring high-quality 4D-CBCT (cone beam computed tomography) which captures these anatomical changes is crucial for successful IGRT (image-guided radiothephy). Additionally, the ability to maintain sufficient anatomical detail is also a key issue.
Methods
Using the public dataset 4D-Lung, which contains planning CT (pCT) and CBCT images of 20 patients with an interval of several days between the acquisition of pCT and CBCT scans. 14 patients were assigned to the training set and 3 to the test set. Rigid and deformable registrations were performed on the CBCT and pCT images. A conditional denoising diffusion model was constructed, which incorporates an edge-constrained boundary(EABC) loss to extract edge information for better capture of internal anatomical details. Visual and quantitative analyses were conducted to evaluate the generated synthetic CT (sCT) images.
Results
The model was evaluated using Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Normalized Cross-Correlation (NCC), and Structural Similarity Index (SSIM). Compared to the ground truth pCT images, the generated sCT images showed a decrease in MAE from 172.38 HU to 53.65 HU, an increase in PSNR from 15.73 dB to 20.41 dB, an improvement in NCC from 0.826 to 0.938, and an improvement in SSIM from 0.783 to 0.927.In the ablation experiments without EABC, MAE decreased from 174.54 HU to 58.77HU, PSNR increased from 15.42 dB to 18.93 dB, NCC improved from 0.803 to 0.915, and SSIM rose from 0.794 to 0.902.
Conclusion
This study constructed a diffusion model trained on registered 4D-CBCT and 4D-CT image pairs. The results demonstrate that the quality of sCT generated by the proposed generative model is superior to that of the registered CBCT.