1167 - Neural Reconstruction for Super-Resolution Radiotherapy QA from Sparse Dosimeter Arrays
Presenter(s)
S. Wang1, C. T. Gibson2, M. R. M. Ashraf3, G. A. Szalkowski4, L. Xing1, and L. Wang5; 1Department of Radiation Oncology, Stanford University, Stanford, CA, 2Stanford Health Care, Palo Alto, CA, 3Department of Radiation Oncology, Stanford University, Palo Alto, CA, 4Georgia Institute of Technology, Atlanta, GA, 5Department of Radiation Oncology, Stanford University School of Medicine, Palo Alto, CA
Purpose/Objective(s): Study presents a neural network-based super-resolution framework for enhancing resolution of sparse dosimetry measurements in patient-specific quality assurance (QA) for radiotherapy. Commercially available QA devices, such as MapCHECK (7 mm spacing) and patient-specific quality assurance (PSQA) (10 mm spacing), rely on sparse detector arrays due to manufacturing complexity and cost. This sparse sampling creates uncertainty in assessing neighboring regions and resolving fine-scale dose variations, especially in high-gradient areas critical for treatment accuracy. Gamma analysis is limited to individual detector points, as interpolation methods fail to provide reliable full-distribution comparisons. Proposed framework reconstructs high-resolution dose distributions from sparse measurements, enabling comprehensive gamma analysis across full dose map with high accuracy.
Materials/Methods: The proposed framework utilizes an implicit neural representation (INR) model trained on paired treatment planning data and corresponding PSQA measurements. During QA, the model processes a single PSQA measurement and the corresponding treatment plan to generate a high-resolution dose distribution. The workflow involves using sparse PSQA data as input and leveraging the learned spatial relationships from the treatment plan to predict the complete dose map at sub-millimeter resolution. To ensure robustness of measurement prediction, 5 mm shifted PSQA measurements were used as independent test data, assessing the model’s ability to handle real-world deviations while maintaining spatial fidelity.
Results: The INR framework enables full-map gamma analysis unavailable with traditional point-based methods. Unlike current commercial solutions that restrict analysis to detector points due to sparse measurement, our framework enhances the spatial accuracy from 10 mm to 1 mm resolution. Validation tests demonstrated 100% alignment between the super-resolution dose reconstruction and real-world measurements at 1 mm/1% gamma criteria, confirming the model’s ability to precisely replicate measured dose distributions. Case studies across different site planning are conducted to demonstrate how super-resolution measurement can reveal additional details hidden by sparse measurement.
Conclusion: This study presents an easy and reliable high-resolution dosimetry solution designed for routine radiotherapy QA. By reconstructing detailed dose distributions from sparse measurements, the proposed framework enables comprehensive full-map gamma analysis, providing spatial information about dose delivery accuracy beyond traditional point-based validation. The demonstrated 100% alignment between reconstructed and real-world measurements validates the model's clinical applicability. This approach enhances treatment validation capabilities without requiring additional measurement time, supporting efficient integration into clinical workflows.