ESTRO 2020 Abstract Book

S115 ESTRO 2020

Material and Methods We gathered 146 abdominal CTs of patients aged 1 to 8 years by involving 5 medical institutes world-wide, and delineated 4 Organs-At-Risk (OARs): liver, spleen, kidneys, and spinal cord. We generated 300 historical-like plans for kidney cancer: renal flank irradiation by AP-PA beams, optionally blocking part of liver or spleen, and with prescribed dose of 14.4 Gy. We set up a pipeline (Fig. 1) to simulate the dose for each plan on each patient automatically, obtaining 43,800 dose reconstructions. For each OAR, we collected mean dose (Dmean), D2cc, V5Gy and V10Gy, as response variables to be predicted by ML. As explanatory variables, we used patient features typically available in historical records (e.g., age, nephrectomy), features detectable by image processing on 2D coronal digitally reconstructed radiographs (DRRs) (e.g., rib cage width, spine length), and features of the RT plan (e.g., field size). The resulting dataset was processed with an evolutionary ML algorithm (GP-GOMEA). A model was learned for each dose-volume metric-OAR pair. Cross- validation was performed by training on 4/5th of the data and validating on 1/5th. This was repeated 50 times to account for stochasticity. Results Table 1 shows the mean absolute errors between ML predictions and ground-truth. The largest errors are obtained for OARs at the edge of the RT field, where dose gradients are steep. These are the contralateral kidneys (2.3/1.9 Gy error for D2cc of left/right kidney), liver for right-sided plans (1.5 Gy error for Dmean), and spleen for left-sided plans (1.6 Gy error for D2cc). The spinal cord is positioned within the field for all plans, leading to uniformly high doses and small prediction errors. For liver and spleen where Dmean deviates the most, our results compare favorably with respect to previous work using age and gender-matched surrogate CTs (same age group): 1.1 vs 1.6 Gy for liver, 1.2 vs 2.6 Gy for spleen. For all OARs and metrics, ML predictions have no systematic bias (mean deviation ~0). Conclusion We developed and experimentally validated a novel approach to dose reconstruction for pediatric abdominal RT that leverages ML. Models relating dose-volume metrics to RT plan, 2D images, and patient features were automatically learned that have unbiased and small errors. We will use these models on actual historical data for dose- risk modeling, aimed at ultimately improving pediatric cancer treatment.

Conclusion Fully automated heart dose calculation from cine MV images of tangential breast fields was proposed, developed and shown to be highly accurate. It may be used in a surveillance program with automated heart dose monitoring of all breast cancer treatments in a clinic. OC-0225 Highly-individualized dose reconstruction for pediatric abdominal radiotherapy with machine learning M. Virgolin 1 , Z. Wang 2 , B.V. Balgobind 2 , I.W.E.M. Van Dijk 2 , J. Wiersma 2 , D.C. Hodgson 3 , A. Bryce-Atkinson 4 , M. Van Herk 5 , C.R.N. Rasch 6 , L. Zadravec Zaletel 7 , P.S. Kroon 8 , G.O. Janssens 8 , A. Bel 2 , P.A.N. Bosman 1,9 , T. Alderliesten 2 1 Centrum Wiskunde & Informatica, Life Sciences and Health Group, Amsterdam, The Netherlands ; 2 Amsterdam UMC-University of Amsterdam, Department of Radiation Oncology, Amsterdam, The Netherlands ; 3 Princess Margaret Cancer Centre, Department of Radiation Oncology, Toronto, Canada ; 4 University of Manchester, Faculty of Biology- Medicine and Health, Manchester, United Kingdom ; 5 University of Manchester, Department of Cancer Sciences, Manchester, United Kingdom ; 6 Leiden University Medical Center, Department of Radiation Oncology, Leiden, The Netherlands ; 7 Institute of Oncology Ljubljana, Department of Radiation Oncology, Ljubljana, Slovenia ; 8 University Medical Center Utrecht, Department of Radiotherapy, Utrecht, The Netherlands ; 9 Delft University of Technology, Department of Software Technology, Delft, The Netherlands Purpose or Objective Knowledge of 3D dose distributions in the anatomy of pediatric cancer survivors is crucial to understand how radiotherapy (RT) is related to late adverse effects. However for patients treated in the past only 2D radiographs were taken. To compensate for this, current methods for dose reconstruction use 3D surrogate anatomies, but discrepancies with real patient anatomies can lead to inaccuracies. We aim to improve dose reconstruction by proposing a novel approach that, instead of using surrogates, employs Machine Learning (ML) to predict dose-volume metrics.