PT - JOURNAL ARTICLE AU - Mathieu Hatt AU - Marie-Charlotte Desseroit AU - Florent Tixier AU - Dimitris Visvikis AU - Catherine Cheze Le Rest TI - Prospective validation of Support Vector Machine based prognostic models exploiting combined PET and CT Radiomics features for Non-Small Cell Lung Cancer: impact of image discretization in texture calculations DP - 2017 May 01 TA - Journal of Nuclear Medicine PG - 721--721 VI - 58 IP - supplement 1 4099 - http://jnm.snmjournals.org/content/58/supplement_1/721.short 4100 - http://jnm.snmjournals.org/content/58/supplement_1/721.full SO - J Nucl Med2017 May 01; 58 AB - 721Objectives: Our goal was to further investigate the validation of prognostic models exploiting FDG PET/CT radiomics features of NSCLC primary tumors and built through machine learning methods, with a focus on the impact of the pre-processing discretization choicesMethods: 116 stage I-III NSCLC patients treated between 2008 and 2012 and retrospectively included constituted the training cohort. Primary tumor volumes were automatically delineated on PET and low-dose CT using the FLAB algorithm and 3D-slicer respectively. Shape, intensity and texture features previously identified as reliable were used to quantify FDG PET and low-dose CT tumors characteristics. For texture analysis, three different discretization methods were considered: linear, fixed bin width and histogram equalization. Support vector machines (SVM) with recursive feature extraction and 10-fold cross-validation were used to combine the most valuable and least redundant features amongst Radiomics and clinical variables. Primary endpoint was survival <6 months. Four different models were built, depending on the quantization method; the 4th model was obtained when using all 3 sets of features. These models were then tested in a prospectively recruited cohort of 55 patients treated between 2014 and 2016.Results: All 4 models built in the retrospective training cohort allowed the identification of the high risk group with poor prognosis (median survival of 4.3 to 6.7 months depending on the model, 3-year survival ≤13%) with accuracy 80-85% by combining 5 variables (PET/CT radiomics and clinical parameters). On the other hand, only two models (with linear discretization and with all 3 sets of features) had satisfactory performance when applied to the prospective validation cohort. Similar results were observed with different endpoints (e.g. survival above or below the median survival of 18 months).Conclusion: Our results highlight the prime importance of the pre-processing discretization step when exploiting textural features for building PET/CT multiparametric prognostic models, as well as the requirement of validating the trained/learned models obtained through machine learning in an independent prospective validation cohort to ensure the robustness of the results. Research Support: With funding of VARIAN, Brest Metropole Oceane, and INCa (project PRINCE #R16063NN)