Purpose: Conventional wrapper and filter-based feature selection methods in radiomics study analyze the correlation of features and endpoints sequentially or individually, often resulting in excluding predictive features when combining with others. In this study, we developed an embedded feature selection method for an artificial neural network (ANN) using Group-Lasso regularization to predict obesity of pediatric patients after proton therapy using magnetic resonance imaging (MRI) features.
Methods: Eighty-four children diagnosed with craniopharyngioma (aged 1-20 years) and treated with surgery and proton therapy were included in this study. Consent forms with Institutional review board approval was obtained. The endpoints were defined as obesity in patients with the normalized body-mass index (BMI) crossing 95% after 5-year follow-up. Imaging features were retrospectively calculated from pre-operative MRI images with T1-, T2-weighted, and FLAIR sequences using LIFEx (Orsay, France) software. Group-Lasso regularization was implemented in a 4-layer ANN developed with TensorFlow (Google, USA) for feature selection. The optimal combination of features was selected by thresholding their L2-norm of weights. Five out of 105 features were selected with the embedded method. Data augmentation was used to reduce bias of the unbalanced grouping and 5-fold cross-validation was used to reduce statistical bias of the ANN model.
Results: The average accuracy of the testing dataset in the 5-fold cross validation was 78.3%±7.8%, 76.7%±9.7%, and 80.0%±7.7% for T1, T2, and FLAIR images, respectively. The macro average area under the ROC curve (AUC) was 0.79, 0.84, and 0.92 for T1, T2, and FLAIR images, respectively.
Conclusion: An embedded feature selection method was implemented with the Group-Lasso regularization for an ANN radiomics model. The dimension of feature space was reduced without compromising average accuracy and AUC. The prediction of obesity could be applied in a personalized treatment plan to improve the quality of post-treatment life.
Proton NMR, Feature Selection, Modeling