ESTRO 37 Abstract book

ESTRO 37

S486

Conclusion iDR reduces range uncertainties related to anatomical variations with high precision and accuracy in the whole patient and hence has the potential to by-pass the necessity of physician’s approval of a ‘daily-plan’, as long as dose stability can be assumed. Motion (as well as setup and range errors) can be taken into account by including robust optimization in the dose restoration. Restoring clinically-approved dose distribution on repeated CTs does not require new ROI segmentation and is therefore compatible with an online adaptive workflow. Ensuring dose stability will be sufficient for the majority of clinical cases. Large changes of relative position of OAR and target volumes are not addressed here and should be taken into account by offline adaptations. PO-0908 Developing Whole Breast Radiotherapy Automatic-Planning System using Beamlet Feature based Model Y. Sheng 1 , T. Li 2 , S. Yoo 1 , F.F. Yin 1 , R. Blitzblau 1 , J. Horton 1 , M. Palta 1 , Y. Ge 3 , Q.J. Wu 1 1 Duke University Medical Center, Radiation Oncology, Durham, USA 2 Thomas Jefferson Univeristy, Radiation Oncology, Philadelphia, USA 3 University of North Carolina at Charlotte, Collage of Computing and Informatics, Charlotte, USA Purpose or Objective High quality treatment planning for whole breast radiation therapy (RT) using tangential fields currently requires manual fluence editing that may take 1-4 hours. This study aims at enabling automatic planning using knowledge models. Material and Methods Whole breast RT plans from 20 patients at Institution A treated with single energy (SE, 6MV, 10 patients) or mixed energy (ME, 6/15MV, 10 patients) were randomly selected for model training. The planning process for a new case consists of three fully automated steps: 1. Energy Selection. To build the model, principle component analysis (PCA) utilized the digital reconstructed radiograph (DRR) of the training cases to extract anatomy-energy relationship and features. A binary classification model based on DRRs gray level automatically selects energy for treatment planning. 2. Fluence Estimation. A random forest (RF) model based on beamlet intensity features generates the initial fluence. 3. Fluence Fine-tuning. This step equalizes the overall dose contribution throughout the whole breast from medial and lateral beam by automatically selecting reference points and applying centrality correction to move reference points to the half penetration depth. The proposed method was validated by comparing automatic plans with manually generated clinical plans using Wilcoxon Signed-Rank test. Dosimetric evaluation included target volume coverage (breast clinical target volume), hotspot volume and lung dose. The test cohort included 31 patients (20 cases from Institution A and 11 cases from Institution B) to cross evaluate the performance of the system. Results For intra-institution validation, in 19/20 cases the model suggested the same energy option as in clinical plans. The target volume coverage (V100%) was 78.1±4.7% for automatic plans, and 79.3±4.8% for clinical plans (p=0.12). Volumes receiving 105% Rx were consistently smaller for automatic plans, (69.2±78.0cc vs 83.9±87.2cc)

although not statistically significant (p=0.13). The mean V10Gy and V20Gy of the lung was 24.4±6.7% vs 24.6±6.7% and 18.6±6.0% vs. 18.9±6.1% for automatic plans and clinical plans (p=0.04, <0.001). The cross institution validation cases included different breast size (irradiated volumes defined by jaw/MLC opening range from 875cc to 3516cc), and the clinical plans used field-in-field technique. For cross institution validation, in 9 out of 11 cases single/mixed energy choice made by the system agreed with clinical plans. V100% were similar (p=0.223) between automatic plans and clinical plans (57.6±8.9% vs. 54.8±9.5%). The absolute V105% was reduced from 395.6±359.9cc in clinical plans to 108.7±163cc in automatic plans (p =0.001). All optimizations were finished within 1.5min. Conclusion Cross institution study results revealed that tangential field treatment showed improved hotspot volume over field-in-field treatment while target coverage was comparable. The system generates breast radiotherapy treatment plans with accurate energy selection, similar target volume coverage, reduced hotspot volumes, and significant reduction in planning time compared to manual planning. N. Wahl 1,2,3 , P. Hennig 4 , H.P. Wieser 1,3,5 , M. Bangert 1,3 1 German Cancer Research Center DKFZ, Medical Physics in Radiation Oncology, Heidelberg, Germany 2 University of Heidelberg, Department of physics and astronomy, Heidelberg, Germany 3 Heidelberg Institute of Radiation Oncology, HIRO, Heidelberg, Germany 4 Max Planck Institute for Intelligent Systems, Probabilistic Numerics, Tübingen, Germany 5 University of Heidelberg, Medical Faculty, Heidelberg, Germany Purpose or Objective Uncertainties play a critical role in radiotherapy planning, especially for charged particles. However, in clinical practice, decisions are based on simulations of the nominal treatment scenario. Computation of the expected treatment and error estimates is restricted to academic studies in uncertainty mitigation, usually based on worst-case or otherwise sampled scenarios. Yet, those sampling approaches imply computational (due to multiple dose calculations) and conceptual (i.e. in fractionated treatments) limitations, especially when applied in optimization. Analytical probabilistic modeling (APM) for radiotherapy planning allows radically new approaches for uncertainty management through closed- form computation of expected dose E[ d ] and its (co)variance Σ d . Here, we extend APM to uncertainty quantification in dose quality metrics/objectives. Material and Methods We consider two analytical methods for propagating uncertainties through a metric/objective function I ( d ): (1) computing the moments by integrating I ( d ) against the probability density over d , and (2) approximating E[ I ( d )] and Var[ I ( d )] through multivariate Taylor expansion on I ( d ). Both formalisms are applied on the functions for mean dose, DVH points, min/max dose d min/max , EUD k , and the piece-wise squared dose objective F +/- . For our derivations the dose is assumed to follow a multivariate normal distribution. E[ d ] and Σ d are estimated from 100 treatment samples for 1 and 30 fractions. We model a PO-0909 Analytical probabilistic models for dose quality metrics and optimization objectives