S2329
Physics - Quality assurance and auditing
ESTRO 2026
Material/Methods: A total of 23 HyperArc plans were generated on a Varian TrueBeam HD accelerator, each corresponding to a single lesion ranging from 0.5 to 32.7 cc. First, an unrestricted plan was created, incorporating only maximum dose constraints and the Normal Tissue Objective (NTO). Additional plans were then generated using the ASC (set to “Very High”) and with MU restricted to 50% of the value in the unrestricted plan, while keeping all other optimization parameters unchanged. Quality assurance (QA) measurements were performed with PTW Octavius4D and a Semiflex 3D ionization chamber. Complexity was calculated using the formula proposed by Younge et al.(1), and dedicated software was developed for simultaneous visualization of MLC apertures and complexity in the three plan types.
that plaomplexity can reliably forecast QA performance in HyperArc treatments. References:
1- Predicting deliverability of volumetric-modulated arc therapy (VMAT) plans using aperture complexity analysis. Kelly C. Younge, Don Roberts, Lindsay A. Janes, Carlos Anderson, Jean M. Moran, Martha M. Matuszak. https://doi.org/10.1120/jacmp.v17i4.6241. Keywords: Hyperarc, Complexity
Proffered Paper 559
Prediction and interpretation of gamma map failures in breast IMRT using deep learning and explainability Eva Ambroa 1 , Pedro Gallego 2 , Jaime Pérez-Alija 2 , Manuel De La Cruz 1 , Oscar Pera 1 , Jaime Quera 1 , Marti Lacruz 1 , Enric Fernández-Velilla 1 , Oliver Díaz 3 , Simone Balocco 3 1 Física i Protecció Radiològica, Hospital del Mar, Barcelona, Spain. 2 Radiofísica i Radioprotecció, Hospital de la Santa Creu i Sant Pau, Barcelona, Spain. 3 Departament de Matemàtiques i informática, Universitat de Barcelona, Barcelona, Spain Purpose/Objective: Traditional studies on gamma passing rate (GPR) prediction focus on classifying treatment fields as pass/fail, but do not address why or where discrepancies occur. This work presents a novel deep learning framework to predict full gamma maps in breast IMRT portal verification, enabling detailed analysis of discrepancies and their underlying causes through explainability tools. Material/Methods: A total of 4646 IMRT fields from 645 patients were analyzed. The dataset was divided into three subsets: training (80%), validation (10%), and testing (10%), using a GPR criterion of 3%/2mm > 90%.A CNN multimodal model was developed to predict the two- dimensional gamma map using as input: predicted portal dose images, MLC modulation maps (MM), and plan complexity metrics. Additionally, a custom loss function was designed to assign higher weight to negative gamma values (around 8%), thereby improving the model’s sensitivity.Explainability was achieved through SHAP applied at pixel level. Regions of poor agreement were further analyzed by aggregating SHAP values to identify the factors that mostly induce the discrepancy. Discrepancies were categorized into three error types (high fluence modulation, MLC speed and mixed causes), each associated with proposed corrective actions (fluence smoothing, dose rate reduction, or replanning). Results: The model successfully predicted full gamma
Results:
Although all individual results were within tolerance, the FREE and ASC plans exhibited greater dispersion compared with the MU-limited plans. In Octavius4D measurements, all but one MU plan exceeded 90% for the gamma criterion (2%, 1 mm), while most FREE and ASC plans scored below 85%. Conclusion: Plan complexity can serve as a predictor of patient- specific QA outcomes without compromising dosimetry. Both point measurements and Octavius4D analyses demonstrated greater stability when complexity was reduced. Conversely, medium- or high- complexity plans showed wide variability, confirming
Made with FlippingBook - Share PDF online