TY - JOUR T1 - [<sup>18</sup>F] FDG PET radiomics to predict disease free survival in Cervical Cancer: a multi-scanner/center study with external validation JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1406 LP - 1406 VL - 62 IS - supplement 1 AU - Marta Ferreira AU - Pierre Lovinfosse AU - Johanne Hermesse AU - Marjolein Decuypere AU - Caroline Rousseau AU - Francois Lucia AU - Ulrike Schick AU - Caroline Reinhold AU - Philippe Robin AU - Mathieu Hatt AU - Dimitris Visvikis AU - Claire Bernard AU - Ralph T.H. Leijenaar AU - Frederic Kridelka AU - Philippe Lambin AU - Patrick E. Meyer AU - Roland Hustinx Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/1406.abstract N2 - 1406Objectives: To test the performances of native and tumour to liver ratio (TLR) radiomic features extracted from pre-treatment 2-[18F] fluoro-2-deoxy-D-glucose ([18F]FDG) PET/CT and combined with machine learning (ML) for predicting disease free survival (DFS) in patients with locally advanced cervical cancer (LACC). Methods: One hundred fifty-eight patients with LACC from multiple centers were retrospectively included in the study. Tumours were segmented using the Fuzzy Local Adaptive Bayesian (FLAB) algorithm. Radiomic features were extracted from the tumours and from regions drawn over the normal liver. We evaluated the predictive value of native radiomics features and tumour to liver ratios (TLR) radiomic features. We separated our data into training and testing sets (80% and 20% of the data from each scanner, respectively). First, a cox proportional hazard model was used to test statistical significance of clinical, treatment and radiomic features. Afterwards, we evaluated whether combining different feature selection (FS) and machine learning (ML) classifiers methods was able to find a radiomics signature to predict DFS. For that purpose we tested a different set of models, which differ in i) the features type, i.e., native or TLR radiomics, ii) the pre-processing of the PET images, i.e., with or without interpolation, iii) the pre-processing of the features, i.e., with or without ComBat harmonization and intensity discretization scheme (fixed bin width or fixed bin number (FBN)), iv) the FS and ML classifier method and v) the metric used to optimize the number of features used in the model. Five-fold cross validation was used to tune the number of features. The models with the selected features were trained using bootstrapping and tested in data from each scanner independently. We considered as best model the one with higher bootstrap F1-score. Reproducibility of radiomics features, clinical data added value and effect of ComBat-based harmonization after FS were evaluated across scanners. The radiomics pipeline used is shown in figure 1. Results: After a median follow up of 28 months, 29% of the patients recurred. No individual treatment scheme, radiomic or clinical features were significantly associated with cancer recurrence. The best model was obtained using 10 TLR features combined with clinical information. Features were selected with the Forward maximum relevance minimum redundancy method based on the mutual information, discretized with FBN (32 bins) and classified with random forest. The area under the curve (AUC), F1 -score, precision and recall were respectively 0.78, 0.49, 0.42 and 0.63. Kaplan-Meier curve of the model was significantly discriminant: log-rank P-value of 0.002(figure 2). Neither using the ComBat-harmonized features nor applying ComBat harmonization on these two models did improve performances. Both the TLR and the native models performance varied across scanners used in the test set (table 1). Conclusions: [18F]FDG PET radiomic features combined with ML add relevant information to the standard clinical parameters in terms of LACC patient’s outcome but remain subject to variability across PET/CT devices. View this table:Predictive performance of best TLR model in each distinct test set ER -