S1870
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
Digital Poster 2037
Quantifying algorithm-dependent dose variations for small lung tumors treated with stereotactic body radiation therapy Eric Grönlund, Marcus Krantz, Thomas Henry, Roumiana Chakarova Department of Medical Physics and Biomedical Engineering, Sahlgrenska University Hospital, Gothenburg, Sweden Purpose/Objective: To quantify variations in dose determination for small lung tumors when using different dose calculation algorithms. Material/Methods: Fifteen clinical lung SBRT VMAT plans using 6FFF photon energy were recalculated with four calculation algorithms and two dose reporting methods (dose-to- water, Dw, or dose-to-medium, Dm) (Table 1). The DVH metrics D1%, D2%, D50%, D98% and D99% were evaluated for the PTV in each calculation and compared against an in-house Monte Carlo reference.Table 1. Dose calculation algorithms from different vendors included in the study.VendorCalculation algorithmDwDm Eclipse (Varian Medical Systems)Analytical Anisotropic Algorithm (AAA)xAcuros XB (AXB)xxRayStation (RaySearch Laboratories AB) (RS)Collapsed Cone (CC)xMonte Carlo (MC)xDoseCheck™ (Sun Nuclear Corporation) (DC)Collapsed Cone (CC)xIn-house Monte Carlo from EGSnrc code packageMonte Carlo (MC)xx Results: For Dw reporting (Figure 1), RS CC systematically overestimated all DVH metrics by an average of 3.2– 3.9 percentage points (p.p.). For D50%, D98%, and D99%, AAA also showed systematic overestimation (3.4–6.4 p.p.), whereas AXB remained within 1.0 p.p. for the same metrics. However, both AAA and AXB underestimated high-dose metrics (D1% and D2%) by an average of –2.6 p.p.For Dm reporting (Figure 2), DC CC underestimated all DVH metrics compared to the in-house MC, with average differences ranging from – 5.8 p.p. (D99%) to –6.7 p.p. (D50%) and –6.6 p.p. (D1% and D2%). Moreover, AXB agreed closely with the in- house MC for D98% and D99% (on average within 0.8 p.p.) and differed by –1.4 p.p. for D50%; however, AXB underestimated D1% and D2% (on average by –3.4 to – 3.2 p.p.). RS MC remained on average within 1.1 p.p. of the in-house MC for all DVH metrics.The median PTV volume was 15.0 cm ³ (range 8.4–106.1 cm ³ ). For PTVs smaller than the median, the observed differences between algorithms became even more pronounced.
Conclusion: The proposed MATE-UNet framework was successfully validated, achieving all clinical goals on 100% of the testing plans. This work demonstrated MATE-UNet's potential as a core component for a fully automated, anatomy-to-plan pipeline for VMAT in pancreatic cancer. References: 1. Çiçek, Özgün, et al. "3D U-Net: learning dense volumetric segmentation from sparse annotation." International conference on medical image computing and computer-assisted intervention. Cham: Springer International Publishing, 2016.2. Oktay, Ozan, et al. "Attention U-Net: Learning where to look for the pancreas." arXiv preprint arXiv:1804.03999 (2018).3. Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017). Keywords: Auto-planning, Deep learning, Pancreatic cancer
Made with FlippingBook - Share PDF online