ESTRO 38 Abstract book

S1039 ESTRO 38

Conclusion Random Forests performed well in the prediction of DR and DD with Radiomics features. The nature of Random Forests and the use of OOB prediction prevents for overfitting. Also the low standard deviation of AUC values when repeating the training are encouraging. Nevertheless, patient population needs to be enlarged in order to be able to perform validation in an independent dataset. EP-1913 Distributed rapid learning made easy: a user- friendly dashboard for model development and execution J. Van Soest 1 , C. Masciocchi 2 , P. Fick 1 , T. Hendriks 1 , R. Negro 2 , A. Damiani 2 , V. Valentini 2 , A. Dekker 1 1 Maastricht Radiation Oncology Maastro Clinic, Knowledge Engineering, Maastricht, The Netherlands ; 2 Università Cattolica del Sacro Cuore, Radiotherapy, Rome, Italy Purpose or Objective Data analysis is becoming more apparent in radiotherapy and medicine in general. Next to the technical skills, interpretation and clinical usefulness of results is of major importance. Hence, visual representations of data and outcomes (plots, charts, nomograms) can help in exploring the data and analysis results. However, the number of people managing both technical and clinical skills is limited, introducing a challenge to leverage from all available (and continuously updated) data. The aim of this abstract is to develop a user-friendly dashboard to ask different types of questions (e.g. showing basic cohort statistics or learning a prediction model). This dashboard triggers these tasks on multiple routine clinical datasets; performing a distributed and privacy-preserving analysis without requiring in-depth technical skills. Material and Methods The datasets available in participating institutes contained 2469 and 837 routine clinical rectal cancer patients. Information regarding age, gender, clinical TNM stage, prescribed RT dose, overall survival status and time were available for inclusion and/or analysis. Three distributed algorithms were developed to perform: a) general statistics, b) plotting one integrated Kaplan- Meier graph, c) learning a Cox proportional hazards model. We developed a dashboard to select the distributed algorithm (A, B or C) and define inclusion criteria. The dashboard also shows previous executions of algorithms (including selected inclusion criteria) and visualizes the result (tables/figures/text) for every algorithm execution. When Algorithm C is selected, the dashboard can show an interactive prediction model executor. In this model executor, users can enter patient-specific criteria whereafter the predicted probabilities over time are The developed dashboard is publicly available at http://coraldashboard.jvsoest.eu. Results of algorithms are visualized (e.g. variable distributions, KM curve for inclusion criteria, prediction model), and stored in the history of algorithm executions. This means users can shown. Results

EP-1912 Outcome prediction with CT radiomics and random forests in primary lung tumor treated with SBRT C. Martín 1 , D. Sevillano 2 , C. Vallejo 3 , M. Martín 3 , J. .D. García 2 , R. Colmenares 2 , R. Morís 2 , B. Capuz 2 , M. Cámara 2 , A. Martínez 2 , F. Orozco 2 , M.J. Béjar 2 , D. Prieto 2 , S. Sancho 3 , F. García-Vicente 2 1 ETSIT. Universidad Politécnica de Madrid, Biomedial Engineering, Madrid, Spain ; 2 Hospital Universitario Ramón y Cajal, Medical Physics, Madrid, Spain ; 3 Hospital Universitario Ramón y Cajal, Radiation Oncology, Madrid, Spain Purpose or Objective To predict treatment outcomes from radiomics information of tumor images in the planning CT. Material and Methods 87 patients treated from September 2012 to December 2016 (median follow-up time=25.2 months) were retrospectively analyzed for this work. Radiomic features from the tumor contoured by the radiation oncologist in the planning CT were obtained with IBEX software ( Zhang et al, IBEX: an open infrastructure software platform to facilitate collaborative work in radiomics, Med Phys 2015 (42)3, 1341-53 ). From there, 766 features regarding tumor shape, intensity levels, Co-occurrence matrix, gradient orientations and intensity difference between neighboring voxels were obtained. Analyzed treatment outcomes were Distant recurrence (DR, 26 cases) and death from disease (DD, 23 cases) For feature selection, a Random Forests algorithm was used with the whole dataset features. Importance of each feature was obtained by permuting the values of each variable and calculating the variation in Out-of-Bag (OOB) error prediction. This process was repeated ten times and mean error increase calculated. Those features with higher mean error increase are those with more prediction power. Once features are selected, Random Forests with only those features was used for prediction. Area under the ROC curves (AUC) were obtained from OOB predictions of the algorithm Kaplan-Meier (K-M) plots were obtained by dividing our patient population in half according to median score of the random forests algorithm. Logrank test was performed in order to check if there was a significant difference between both patient populations. Results For DR, two features regarding Co-occurrence Matrix maximum probability and local standard deviation of intensity values were selected. The algorithm achieved an AUC of 0.76 (SD= 0.02). Also, a value of p=3.10 -5 was obtained in the logrank test. In the case of DD, features selected were related to information measure of the Co-occurrence matrix and maximum probability in co-ocuurrence matrix. The value of AUC was 0.79 (SD= 0.01) and logrank test yielded a value of p=0.03. K-M plots are shown in Figure 1.