Author: David J. Carlson, Huixiao Chen, Tianqi Chen, Jun Hou, Chi Liu, Qiong Liu, Henry S. Park, Huidong Xie 👨🔬
Affiliation: Yale University, Department of Therapeutic Radiology, Yale University School of Medicine 🌍
Purpose:
The RefleXion® X1 Biology-guided radiotherapy (BgRT) system consists of dual PET detectors, a 6MV linear accelerator (linac), a 64-leaf collimator, an MVD detector, and a CT scanner mounted on a gantry that rotates at 60 rpm. The PET signal from the tumor is detected by the PET system and used to guide the radiation beam from the linac onto the tumor. The BgRT patient workflow involves acquiring planning PET data on the X1 system to generate BgRT treatment plans. This work aims to evaluate whether diffusion-based algorithms can be implemented to improve the RefleXion X1 PET data.
Methods:
We adapted our previously introduced DDPET-3D as our AI model in this study. It is a diffusion model designed for 3D low-dose PET imaging. In our previous work, the model was trained on a large-scale low-count FDG PET dataset and demonstrated superior performance in a multi-center evaluation. The pre-trained DDPET-3D model was directly applied to improve RefleXion X1 PET images for evaluation without further network fine-tuning. To generate paired noisy/clean data for model evaluation, we used an inbuilt third-party PET conversion software in the RefleXion system that can convert clean diagnostic PET data to RefleXion X1-equivalent PET data (Simulated RefleXion PET). For validation, we used four cervical cancer patient studies.
Results:
As shown in the supporting data, clinically relevant volume of the cervical tumor is obscured due to high image artifacts in the X1-equivalent PET data. DDPET-3D effectively reduced image noise and enhanced the quality of the X1-equivalent PET image. The missed cervical tumor region was effectively recovered by the DDPET-3D model. DDPET-3D improved SSIM measurements from 0.911 to 0.960 and reduced the RMSE measurements from 0.622 to 0.572.
Conclusion:
We showed that DDPET-3D, a diffusion-based AI model, can enhance the RefleXion PET image quality for cervical cancer patients.