Super-Resolution Dosimetry: A Deep Learning Framework for Radiotherapy QA
Abstract
Purpose
Sparse detector arrays commonly used for patient-specific radiotherapy quality assurance (QA) cannot provide complete spatial dose distribution measurements, leading to uncertainties particularly in high-gradient dose regions.The goal of this project is to develop a super-resolution dosimetry framework capable of reconstructing high-resolution dose distributions from sparse dosimetry measurements for comprehensive dosimetry analysis .
Methods
An implicit neural representation (INR) model with Fourier feature embeddings was developed in two steps: pretraining on the high-resolution treatment plan and fine-tuning on sparse dosimeter data from ArcCHECKTM. To evaluate robustness, some delivery errors were introduced intentionally by modifying the planned dose distribution. These modified distributions were masked at detector positions to emulate realistic sparse measurements. Super-resolution reconstruction was then performed using the trained INR model, which was evaluated on its ability to recover ground truth error patterns.
Results
Across 12 clinical patient plans, our method achieved an average full-map gamma pass rate of 97.1% using 3%/2 mm criteria, compared to 93.0% for conventional point gamma and 57.0% for bilinear interpolation. In intentionally introduced delivery error cases, the INR framework successfully reconstructed structured anomalies such as horizontal dose-insufficient stripes and diagonal measurement shifts, accurately recovering patterns not easily visualized by sparse measurements. Independent verification using shifted detector measurements yielded 100% gamma agreement, confirming reconstruction loyalty. 2D dose distribution discrepancies remained within 0.24% on average, substantially lower than interpolation-based methods.
Conclusion
The proposed super-resolution QA framework not only improves spatial completeness of dose verification but also enables detection of subtle, clinically relevant delivery errors. By combining prior knowledge from the treatment plan with sparse QA data, this method offers a practical and interpretable enhancement to standard QA practices.