S2407
Physics - Quality assurance and auditing
ESTRO 2026
Digital Poster Highlight 4995
Super-Resolution Dosimetry via Implicit Neural Representation for Patient-Specific Radiotherapy QA Siqi Wang, Siqi Ye, Clinton Gibson, Andrew Heider, Lei Wang, Gregory Arthur Szalkowski, Lei Xing, Ramish Ashraf Radiation Oncology, Stanford University, Palo Alto, USA Purpose/Objective: Sparse detector arrays commonly used in patient- specific radiotherapy quality assurance (QA) limit the spatial completeness of dose verification, particularly in high - gradient regions where steep dose fall - offs occur [1,2]. Such limitations hinder accurate evaluation of delivered dose distributions and can obscure clinically relevant delivery errors. To address this, we propose a super - resolution dosimetry framework that reconstructs high - resolution dose maps from sparse QA measurements using an implicit neural representation (INR) model. By leveraging the treatment plan as prior knowledge and fine - tuning with measured detector data, the proposed method aims to overcome sampling limitations without hardware modifications. This work investigates whether a treatment-informed INR framework can enhance dose reconstruction accuracy, improve gamma pass rates, and enable interpretable visualization of spatial discrepancies in complex IMRT
Results: For the single target case, MicroDiamond and A16 measurements in CubeRT agreed within 1% of each other, and within 3.3% and 2.3% of the plan, respectively. For one multi-lesion case, the average agreement over both targets was 1% (between detectors) and 1.8% and 4.0%, for the ion chamber and microdiamond, respectively. For the single target case, MicroDiamond and A16 measurements in CubeRT agreed within 1% of each other, and within 3.3% and 2.3% of the plan, respectively. For one multi- lesion case, the average agreement over both targets was 1% (between detectors) and 1.8% and 4.0%, for the ion chamber and MicroDiamond, respectively. The other multi-lesion plan showed up to 6% disagreement between planned and measured doses in CubeRT, and this plan was rejected. Measurements from the ACDS Level III audit agreed with the plan within 3.0% on average, and within 3.5% of the CubeRT validation measurements, demonstrating high consistency between CubeRT validation measurements and ACDS audit measurements. Conclusion: This cross-validation study demonstrates strong agreement between point dose measurements in the CubeRT phantom and the ACDS Level III audit, supporting the feasibility of further work to establish reciprocal audit recognition among GHG member organizations, particularly for multi-target tests. Keywords: audit, quality assurance, SRS
and VMAT deliveries. Material/Methods:
An INR model with Fourier feature embeddings [3–5] was developed to encode continuous dose distributions from sparse measurements. The network was first pretrained on high - resolution treatment plan doses to learn plan - specific spatial priors, followed by fine - tuning using sparse ArcCHECK™ diode array data (10 mm spacing). A distance - weighted loss emphasized fidelity near detector positions while ensuring smooth global consistency.Twelve clinical IMRT/VMAT patient QA plans were retrospectively
Made with FlippingBook - Share PDF online