Author: Cheng-En Hsieh, Shen-Hao Li, Hsin-Hon Lin, Shu-Wei Wu, An-Ci Yang 👨🔬
Affiliation: Department of Medical Imaging and Radiological Sciences, Chang Gung University, Proton and Radiation Therapy Center, Chang Gung Memorial Hospital, Proton and Radiation Therapy Center, Chang Gung Memorial Hospital Linkou 🌍
Purpose:
The aim of this study is to develop a framework of generating patient-specific phantom tailored for head and neck proton therapy. From these phantoms, digital reference objects based on the phantom are derived by incorporating ground-truth physical parameters including organ geometries, elemental compositions, stopping power ratios (SPR), and relative electron densities (RED).
Methods:
The framework for generating patient-specific phantoms is based on the XCAT (4D Extended Cardiac-Torso) model, with organ-specific physical properties sourced from ICRP Publication 89. Patient CT images were segmented into major organs using nnUNet, while XCAT CT images were generated using our GPU-accelerated CT simulator. The two CT datasets were registered using both rigid and deformable methods, generating displacement vector fields (DVFs). The DVFs were applied to the XCAT model to create patient-specific phantoms. Evaluation metrics such as Structural Similarity Index (SSIM) and Dice Similarity Coefficient (DSC) were employed to assess the fidelity of the generated phantoms.
Results:
To evaluate the feasibility of this method, CT images were collected from ten head and neck cancer patients undergoing proton therapy. The generated phantoms incorporated essential tissue characteristics with anatomical variations for individual organs. Initial similarity analysis between XCAT CT and patient CT images showed a DSC of 0.63 (±0.011) and SSIM of 0.489 (±0.117). After registration, DSC improved to 0.728 (±0.022), while SSIM increased to 0.875 (±0.045) between twin CT and patient CT images. The generated phantoms demonstrated high anatomical fidelity, accurately capturing organ geometries and tissue properties, despite minor deviations in smaller structures.
Conclusion:
This patient-specific phantoms generation framework provides an effective method for generating twin phantoms, tailored to individual patients. These phantoms are valuable for optimization of treatment planning and AI-based applications. Future efforts will be made to incorporate fast registration techniques to further accelerate and refine the phantom generation process.