S1715
Physics - Dose prediction/calculation, optimisation and applications for particle therapy planning
ESTRO 2026
Conclusion: Unoptimal target coverage in terms of both DRBE and LETd can be associated with reduced LC. While the RBE model choice influences the results, the combined DRBE–LETd effect outperformed dose alone in predicting outcome. These findings suggest that incorporating LETd optimization into CIRT planning could enhance LC for ACC. References: [1] Karger, C. P., & Peschke, P. (2017). RBE and related modeling in carbon-ion therapy. Physics in Medicine & Biology, 63(1), 01TR02. https://doi.org/10.1088/1361-6560/aa9102 [2] Yang Y, Patel SH, Bridhikitti J, et al. Exploratory study of seed spots analysis to characterize dose and linear-energy- transfer effect in adverse event initialization of pencil- beam-scanning proton therapy. Med Phys. 2022; 49: 6237–6252.https://doi.org/10.1002/mp.15859 Keywords: Carbon-ion therapy, Adenoid Cystic Carcinoma, RBE Digital Poster 1305 Automated beam angle selection for lung proton therapy using a deep-learning model Robert Kaderka 1 , Bo-Ru Lin 2,3 , Po-Wei Wu 2 , Yan-Cheng Huang 2 , Keng-Chi Liu 2 , Peter Huang 2 , Yi-Chin Ethan Tu 2 , Mariluz De Ornelas 1 , Chang Chang 4 1 Radiation Oncology, University of Miami, Miami, USA. 2 Medical Solutions, Taiwan AI Labs, Taipei, Taiwan. 3 Data Science Degree Program, National Taiwan University and Academia Sinica, Taipei, Taiwan. 4 Radiation Medicine and Applied Sciences, University of California San Diego, La Jolla, USA Purpose/Objective: Intensity-modulated proton therapy (IMPT) reduces integral dose compared to photons. However, proton planning is complex and requires a complex two-step planning process consisting of beam angle selection followed by fluence optimization. Some automated fluence optimization solutions exist, but angle selection remains a manual task. We propose a novel deep-learning (DL) solution automating beam angle selection for lung IMPT. DL-predicted angles are compared to manually chosen ground truth and resulting lung IMPT plans are evaluated dosimetrically. Material/Methods: A retrospective cohort of 60 lung IMPT patients was identified, with 48 cases for model training and 12 for an independent test set.The workflow includes Pre- Processing, Model Training, DL angle prediction, and Post-Processing. The Pre-Processing step converts DICOM data (CT and RT-Structures) into a 3D tensor format. Model training utilized a MedicalNet 3D convolutional neural network backbone, framing
with dose to derive cumulative dose-LETd volume histograms (DLVHs) [2]. Group differences between LC and PD for median DVHs and DLVHs were assessed via Mann–Whitney test ( α =0.01). DLVH points showing the highest discriminative power (AUROC>0.7) were retained as candidate DRBE–LETd (DL) predictors. Kaplan–Meier analyses, based on the identified DL thresholds, compared high- vs. low-exposure groups, and hazard ratios (HRs) for recurrence were estimated using the Cox model to identify the most significant volumetric threshold. Results: Median DVH differences between PD and LC groups were significant only for DMKM at the lower prescription level (4.1 Gy(RBE)/fx) (Figure 1). When LETd was included, median DLVHs showed significant group differences for both RBE models across multiple DL points (Figure 2).The most discriminative DL levels were:DLEM: 65.4 Gy(RBE) at 51 keV/ μ m (AUC = 0.75) – median volume: 59% (LC) vs 40% (PD)DMKM: 57.8 Gy(RBE) at 51 keV/ μ m (AUC = 0.72) – median volume: 61% (LC) vs 50% (PD)A 70% CTV coverage threshold stratified LC for both models (HRLEM = 0.14, HRMKM = 0.23), indicating an 86% and 77% reduction in local recurrence risk for patients with higher DL coverage, respectively.
Made with FlippingBook - Share PDF online