Author: Guangjun Li, Ying Song, Huanan Tang, Tianxiong Wu, Qiuyi Ye, Wei Zhang 👨🔬
Affiliation: West China Second Hospital of Sichuan University, United Imaging Healthcare, West China Hospital of Sichuan University 🌍
Purpose:
To propose a general low-dose reconstruction model for ultra-sparse-view cone-beam CT (CBCT) and evaluate its clinical application in improving image quality and reducing radiation dose for image-guided radiation therapy (IGRT).
Methods:
Planning CT images of head-and-neck patients were paired with the corresponding simulated CBCT projections at under-sampling rates of 1/8, 1/16, 1/32, and 1/64, forming strictly structure-preserved CT-CBCT pairs. These were divided into training, validation, and testing datasets for model construction and evaluation. A hyper-resolution ultra-sparse-view CBCT reconstruction model was developed using a GAN, named as the planning CT-based strictly-structure-preserved neural network (PSSP-NET). Additionally, clinical CBCT datasets were reconstructed using PSSP-NET (P-CBCT) for clinical test. Subjective evaluation was performed by two oncologists, while objective evaluation utilized PSNR, SSIM, and DSC for the mandible, compared with conventional FDK, SART and DIF-Net algorithms. Set-up errors and dose shifts of radiotherapy treatment plans were assessed by rigid registration of P-CBCT to planning CT.
Results:
In the model test, PSSP-NET consistently maintained high image quality even at an extremely low under-sampling rate of 1/64. P-CBCT achieved clinically acceptable image quality across all under-sampling rates, recognized by oncologists as comparable to planning CT, and outperformed other algorithms in both PSNR and SSIM metrics. During the clinical test, P-CBCT at a 1/8 sampling rate achieved equivalence to full-view FDK reconstruction. The set-up errors were limited to 0.50 mm/-0.60° (model test) and 2.50 mm/2.20° (clinical test). Dose discrepancies were predominantly observed at the tissue-air interface.
Conclusion:
In this study, we proposed a general low-dose reconstruction model for sparse-view CBCT at 1/8, 1/16, 1/32, 1/64 under-sampling rates based on theoretical CT-CBCT simulation and DL. The proposed model, validated in head-and-neck cancer patients, demonstrated superior performance over FDK, SART, and DIF-Net. PSSP-NET was general, fast, efficient, and valuable for reducing daily CBCT dose and enhancing image quality.