ESTRO 38 Abstract book

S558 ESTRO 38

well for treatment planning purposes. Since some contours can deviate from the clinical contours, clinical inspection of automated contours is still required. The combination of automated DLD plus model-based automated planning has the potential to increase the efficiency of routine clinical care, and facilitate online adaptive radiotherapy. PO-1011 Calibration and validation of ion stopping power prediction with Philips IQon Spectral CT F. Faller 1 , B. Ackermann 2 , G. Pahn 3 , M. Alber 1 , W. Stiller 4 , A. Mairani 2 1 University Hospital Heidelberg, Department of Radiation Oncology, Heidelberg, Germany ; 2 Heidelberg Ion-Beam Therapy Center HIT, Department of Radiation Oncology, Heidelberg, Germany ; 3 Philips Healthcare, Clinical Scientist CT, Hamburg, Germany ; 4 University Hospital Heidelberg, Diagnostic Radiology, Heidelberg, Germany Purpose or Objective This study presents for the first time an experimental validation of stopping-power ratio (SPR) prediction from Philips IQon Spectral CT data, exploring its benefit for improving radiotherapy (RT) treatment planning, especially in view of the accuracy requirements imposed by proton and heavy ion therapy. For this purpose, we investigated the added value of derived quantitative image data, such as three-dimensional maps of effective atomic number (Z eff ) and electron density relative to water (ED) covering the whole acquisition field-of-view. The ultimate purpose of this study was the derivation of an accurate three-dimensional map of SPR w based on maps of Z eff and ED acquired with spectral CT imaging technique. Material and Methods The accuracy of ED and Z eff by Philips IQon Spectral CT was verified in phantoms with various tissue-equivalent inserts (Gammex 467, CIRS 062M). SPR w values were determined following the Bethe equation and applying different approaches available in literature aiming to convert Z eff maps in ionization potential (I) maps. The impact of spectral CT scanning settings, regarding acquisition and reconstruction parameters, namely tube potential, acquisition dose, and gantry rotation time, on Z eff and ED maps was assessed. The derived SPR w maps values were validated against experimentally determined SPR meas of the same insert. Comparisons against SPR w derived using spectral CT were performed, evaluating the accuracy of the methodology. Results We calibrated and validated three-dimensional maps of Z eff and ED from spectral CT data with phantom measurements (deviation within a few per cent compared to reference data). Using Z eff data from spectral CT, tissue surrogates can be characterized more accurately than using standard HU-calibration. For the tissue surrogates SPR w predicted from spectral CT images was within a mean accuracy of <1 % compared to SPR meas (Figure 1). The experimental validation in defined materials is a mandatory step to show absolute SPR w errors from spectral CT prediction before considering biological tissue samples. Furthermore, a method for efficient integration of SPR w in the Monte Carlo treatment planning framework available at the Heidelberg Ion Beam Therapy center was developed.

the clinical contour) and with deep learning (DLD). No manual editing was performed on the DLD contours. In order to avoid plan optimization bias, treatment plans were created using knowledge-based planning, based on MD and DLD structures (MD-plans, DLD-plans) and the mean dose to the manually delineated OAR structures was compared. Results Generation of deep learning contours takes ~10 seconds for all OARs. The average dose was statistically significantly higher for DLD-plans for the lower larynx, the inferior PCM, and the esophagus (table 1). Average differences were not clinically significant, but differences in some individual cases could be. From the 209 OARs, 28 (13.4%) received an increase of more than 2Gy with the DLD-plans, and 7 (3.3%) a decrease of more than 2Gy. A full overview of the OARs’ Sørensen-Dice coefficients (SDC) and their increase/decrease in Gy (ΔGy) for DLD- plans compared to MD-plans can be seen in figure 1. Low SDCs were generally the result of variability in the MD or the limited amount of training data.

Conclusion Using a training set consisting of only a relatively limited number of OAR delineations without extensive curating, automated deep learning segmentation for head and neck OARs, is fast and for most organs it performs sufficiently