Author: Sam Beddar, Jason Michael Holmes, Daniel G. Robertson, James J. Sohn, Ethan D. Stolen 👨🔬
Affiliation: Department of Radiation Oncology, Mayo Clinic, MD Anderson Cancer Center, Department of Radiation & Cellular Oncology, University of Chicago, Department of Radiation and Cellular Oncology, University of Chicago 🌍
Purpose: Camera-based scintillation dosimetry incorporating large volumes have shown promise for fast and comprehensive evaluation of external beam treatment fields. While some efforts have been made in 3D reconstruction for scintillation dosimetry, the limited number of camera angles (typically 1-4) has posed significant challenges. Although our previous work demonstrated the feasibility of limited-angle reconstruction for proton pencil beams, existing algorithms remain computationally expensive with imperfect results. In this study, we employ a convolutional neural network (CNN) to provide increased 3D reconstruction accuracy at a decreased computational cost.
Methods: Proton dose distributions in a phantom and patient were generated using a treatment planning system. For each pencil beam spot position, three orthogonal projections were computed by integrating the dose along one spatial dimension. These projections were input to a 3D U-Net model configured with four encoding layers and four corresponding decoding layers. The dataset was divided into 80% training, 10% validation, and 10% test subsets. Model performance was evaluated using mean absolute error (MAE) and gamma analysis of the predicted dose reconstruction for both single spot positions and a single patient field.
Results: The CNN-based dose reconstruction demonstrated high accuracy on the test set, with a MAE of cGy. The average difference for the patient dose distribution was 0.044 cGy with a maximum difference of 4.7 cGy. 3D gamma analysis reported 99.47% (1%, 1 mm) passing. The reconstruction required only 13 ms per spot position and 30.0 s for the patient field.
Conclusion: We demonstrated that a CNN-based approach could overcome the limitations of conventional reconstruction methods for 3D scintillation dosimetry, achieving both high accuracy and computational efficiency. This advancement enables real-time 3D dose verification for proton pencil beams using limited-angle projections. Our approach was verified by a high gamma analysis passing rate for a patient field.