ESTRO 2023 - Abstract Book

S134

Saturday 13 May

ESTRO 2023

Considering the cohort median, the observed NTCP differences due to variability in the RBE of protons remained small. However, for some patients, large ∆ NTCP values were observed, especially for OARs of small volume. Accurate modelling of proton RBE is needed to identify these patients before treatment and to consider the reduction of the individual toxicity risk during treatment planning. [1] Eulitz J et al. Phys Med Biol 2019; 64:225020.

[2] Wedenberg M et al. Acta Oncol. 2013; 52:580–588. [3] Dutz A et al. Radiother Oncol. 2021; 160:69-77.

PD-0173 Proton center variations in predicting pulmonary toxicities from proton radiotherapy of lung cancer T. Nano 1 , G. Valdes 1 , J. Scholey 1 , A. Comas-Leon 1 , E. Gennatas 2 , W. Hartsell 3 , J. Zeng 4 , M. Chuong 5 , M. Mishra 6 , L. Rosen 7 , J. Chang 8 , H. Tsai 9 , J. Urbanic 10 , C. Vargas 11 , L. Ungar 12 , E. Eaton 12 , C. Simone 13 1 University of California, San Francisco, Radiation Oncology, San Francisco, USA; 2 University of California, San Francisco, Epidemiology and Biostatistics, San Francisco, USA; 3 Northwestern Medicine Proton Center, Radiation Oncology, Chicago, USA; 4 University of Washington, Cancer Care Alliance Proton Therapy Center, Seattle, USA; 5 Miami Cancer Institute, Radiation Oncology, Miami, USA; 6 University of Maryland, Proton Treatment Center, Baltimore, USA; 7 Willis-Knighton Medical Center, Radiation Oncology, Shreveport, USA; 8 Oklahoma Proton Center, Radiation Oncology, Oklahoma, USA; 9 New Jersey Procure Proton Therapy Center, Radiation Oncology, Somerset, USA; 10 California Protons Therapy Center, Radiation Oncology, San Diego, USA; 11 Mayo Clinic Proton Center, Radiation Oncology, Phoenix, USA; 12 University of Pennsylvania, Computer and Information Science, Philadelphia, USA; 13 New York Proton Center, Radiation Oncology, New York, USA Purpose or Objective To compare center-specific variation in predicting the probability of grade ≥ 2 pneumonitis or dyspnea within 12 months of receiving conventionally fractionated or mildly hypo-fractionated proton beam therapy (PBT) for lung cancer. Materials and Methods Demographic and treatment characteristics (132 features) of patients treated with PBT were used to predict grade ≥ 2 pulmonary toxicities from 7 institutions enrolled in the Proton Collaborative Group prospective clinical trial NCT01255748. A total number of 868 patients were evaluated across all centers and individual center enrollment ranged from 65 – 229 cases. Machine learning algorithms (Logistic Regression and Random Forrest) were trained using cases from each center and tested on intra-cases (from the same center as the training dataset) and inter-case (outside the center of the training dataset). A double 10-fold cross-validation was performed to tune hyperparameters without leak of information. Balanced Accuracy (BA) and Area under the Curve (AUC) were calculated. Confidence intervals were obtained using bootstrap sampling.

Results From the 868 patients studied, 245 (28.2%) had grade ≥ 2 pulmonary toxicity, and individual center probability of grade ≥ 2 pulmonary toxicity ranged from 5.4% to 36.7% as summarized in Table 1A. When we combined demographic with dosimetric features, the best performing model had

an AUC of 0.75±0.02 and BA of 0.67±0.02. As shown in Table 1B, the intra-case variation in BA and AUC ratio relative to overall performance ranged from 0.88 to 1.47 and 0.82 to 1.36 respectively, indicating same-center predictions can be up to 47% higher. Similarly, the inter-case variation of BA and AUC ratio ranged from 0.86 to 1.05 and 0.88 to 1.1 respectively, indicating that same-center prediction improvements do not necessarily translate when making predictions on outside centers. Figure 1 shows BA and AUC performance of models trained from all centers and tested on all centers individually. Conclusion In this large analysis of prospectively enrolled patients with lung cancer treatment with proton therapy, advanced machine learning methods have shown that predicting grade ≥ 2 pneumonitis or dyspnea with models trained and evaluated within the same center do not necessarily translate to outside institutions and robust models are needed to reduce performance variability of models that are deployed clinically. PD-0174 Radiomics-based prediction of local control of brain metastases after resection and radiotherapy J.A. Buchner 1 , F. Kofler 2 , M. Mayinger 3 , S.M. Christ 3 , T.B. Brunner 4 , A. Wittig 5 , B. Menze 2 , C. Zimmer 6 , B. Meyer 7 , M. Guckenberger 3 , N. Andratschke 3 , R.A. El Shafie 8 , J. Debus 8 , S. Rogers 9 , O. Riesterer 9 , K. Schulze 10 , H.J. Feldmann 10 , O. Blanck 11 , C. Zamboglou 12 , K. Ferentinos 13 , R. Wolff 14 , K.A. Eitz 1 , S.E. Combs 1 , D. Bernhardt 1 , B. Wiestler 6 , J.C. Peeken 1 1 Klinikum rechts der Isar, Technical University of Munich, Department of Radiation Oncology, Munich, Germany; 2 Technical University of Munich, Department of Informatics, Munich, Germany; 3 University Hospital of Zurich, University of Zurich, Department of Radiation Oncology, Zurich, Switzerland; 4 University Hospital Magdeburg, Department of Radiation Oncology, Magdeburg, Germany; 5 University Hospital Jena, Friedrich-Schiller University, Department of Radiotherapy and Radiation Oncology, Jena, Germany; 6 Klinikum rechts der Isar, Technical University of Munich,