ESTRO 38 Abstract book

S518 ESTRO 38

(MI)), or Support Vector Machine (SVM) model. The binary classifiers, trained to predict each histological class (AC, SCC, LCC, and NOS), were linear model or deep neural networks (DNN) (ReLU activation, dropout of 0.35, and 2 layers of 30 neurons each). The performance of the machine learning models was assessed by the computation of the ROC curve, the area under the ROC curve (AUC; measure of separability between classes for any threshold) and the accuracy (the fraction of correct predictions) for each histological class. Results Table 1 displays the AUC and Accuracy on training and validation sets using the best machine learning methods having linear or DNN classifiers. Figure 2 shows the ROC curves obtained for tumor grade prediction on validation set. DNN gave superior modeling of the training set whereas linear models gave better prediction on the validation set. Conclusion We demonstrated that machine learning models trained on CT’s radiomic features could potentially predict histological categorization of NSCLC tumors. Further investigations will be done.

overall survival, OS, and the local control, LC. The quality of the models was appraised by means of the concordance index and the area under the curve, AUC. The significant predictors at univariate analysis were included in a multivariate Cox regression model including the uncorrelated significant features. The multivariate model was then verified on the validation group. Results The analysis of clinical and textural data showed 9 predictors significant at univariate analysis for OS, and 4 predicotrs significant for LC. In the multivariate model, only two variables resulted significant predictors for OS: age and Homogeneity_GLCM with p<0.01, with AUC of 0.80 (95%CI: 0.66-0.94) and 0.73 (95%CI: 0.53-0.93) in the training and validation groups, respectively. Again in the multivariate model, two features were retained: short run low grey level emphasis and grey level non-uniformity, with AUC of 0.65 (95%CI: 0.52-0.81) and 0.61 (95%CI: 0.50- 0.78) for the training and validation sets, respectively. In the low risk group, the median OS and LC in the validation group were 14.4 (95%CI: 12.2-21.2) and 28.6 (95%CI: 12.5- not reached) months, while in the high-risk group were 8.6 (95%CI: 7.0-18.0) and 17.5 (95%CI: 7.6-not reached) months, respectively. Conclusion A CT based radiomic signature was identified which correlate with OS and LC after SBRT for pancreatic adenocarcinoma, and allowed to identify low and high-risk groups of patients. PO-0956 Non Invasive Grading of Non-Small Cell Lung Cancer Using Machine Learning Models Based on Radiomics S. Aouadi 1 , R. Hammoud 1 , T. Torfeh 1 , N. Al-Hammadi 1 1 National Center for Cancer Care & Research, Radiation Oncology, Doha, Qatar Purpose or Objective Grading of non-small cell lung cancer (NSCLC) is crucial for appropriate therapy decisions and prediction of prognosis. Traditionally, invasive histological methods are used. We propose non-invasive approaches for grading NSCLC by training machine learning models on radiomics. Material and Methods 259 patients with NSCLC were collected from the dataset [NSCLC-Radiomics] in The Cancer Imaging Archive. For each patient, pretreatment CT scans, manual delineation of the GTV and histological classification (28 adenocarcinoma (AC), 85 squamous cell carcinoma (SCC), 93 large cell carcinoma (LCC), and 53 adenosquamous carcinoma (NOS) cases) were available. The dataset was randomized and divided into training and validation sets of 200 and 59 patients respectively. CT scans were resampled to 1×1×1 mm³ voxels using three- dimensional Lagrangian polygon interpolation and intensities were divided into bins of 10HU from -1000 to 3000HU. 42 radiomic features were extracted from GTV’s VOI, defined on each patient CT, using LIFEx software (v4.00). They were composed of 3 shape, 8 first order, and 31 texture features. Shape features describe the 3D geometric properties of the tumor. The first order features describe tumor intensity distribution. Texture features, which describe the intra-tumor heterogeneity, were extracted from the gray level co-occurrence (6), neighborhood grey-level different (3), grey-level run length (11), and grey-level zone length (11) matrices. Machine learning models are composed of three blocs: preprocessing, selection and classification. Multiple configurations were benchmarked. Features preprocessing methods were QTRG (mapping data to have normal distribution), MINMAX_SC (linear scaling to [0, 1]), or ROBUST_SC (scaling to the interquartile range). The selection methods of the most representative features were based on elastic-net regularization (ELASTIC-NET), univariate statistical tests (Anova-F or mutual information

PO-0957 Radiomics study from the dose-painting multicenter phase III trial on newly diagnosed glioblastoma F. Tensaouti 1,2 , J. Bailleul 1,2 , E. Martin 3 , F. Desmoulin 2 , S. Ken 4 , J. Desrousseaux 1 , L. Vieillevigne 4 , J. Lotterie 2,5 , V. Lubrano 2,6 , I. Catalaa 2,7 , G. Noël 8 , G. Truc 9 , M. Sunyach 10 , M. Charissoux 11 , N. Magné 12 , P. Auberdiac 13 , T. Filleron 3 , P. Peran 2 , E. Cohen-Jonathan Moyal 1,14,15 , A. Laprie 1,2,15 1 Institut Claudius Regaud/Institut Universitaire du Cancer de Toulouse – Oncopôle, Radiation oncology, Toulouse, France ; 2 ToNIC- Toulouse NeuroImaging Center- Université de Toulouse- Inserm- UPS-, Inserm 1214, Toulouse, France ; 3 Institut Claudius Regaud/Institut Universitaire du Cancer de Toulouse – Oncopôle, Biostatistics, Toulouse, France ; 4 Institut Claudius Regaud/Institut Universitaire du Cancer de Toulouse – Oncopôle, Medical Physics, Toulouse, France ; 5 CHU Toulouse, Nuclear Medicine, Toulouse, France ; 6 CHU Toulouse, Neurosurgery, Toulouse, France ; 7 CHU Toulouse, Radiology, Toulouse, France ; 8 Centre Paul Strauss- EA 3430- University of Strasbourg, Radiation Oncology, Strasbourg, France ; 9 Centre Georges-François Leclerc, Radiation Oncology, Dijon, France ; 10 Centre Léon-Bérard-, Radiation oncology, Lyon, France ; 11 Institut du Cancer de Montpellier, Radiation Oncology,

Made with FlippingBook - Online catalogs