Adaptive feature selection in PET scans based on shared information and multi-label learning
Résumé
Feature selection for tumor treatment outcome prediction in PET scans amounts to determine the best predictors in order to classify treatment outcomes for unseen data. Several challenges have to be addressed, notably the dissensus about the most predictive radiomic features and the relatively small-sized and imbalanced training samples. To overcome these issues, we propose a multi-label learning (MLL) problem formulation that we associated to the technique of feature selection with shared information among multiple tasks. In fact, we opt to simultaneously mining the shared knowledge across similar tasks from MRI and fused PET/MRI images in order to sort PET features according to their relevancy. Then, by repeatedly training an SVM model, different feature sets are elected to be informative for each iteration. Thus, by associating these sets with their correspondents in other iterations, the MLL formulation is performed in order to adequately adapt both feature sets and training instances according to their fit with unseen data. The adaptive synthetic sampling technique is adapted to the framework of MLL data balance. After that, the global and local label correlation technique is used for multi-labeling, while taking into consideration the capacity of MLL on recommending the best sets of features and training instances. Finally, an SVM classifier is applied with both selected features and training sub-samples in order to output prediction results. Conducted experiments show that attributing each feature set to a particular set of instances and training the model with balanced subset of the whole training data allows to outperform relevant methods from the state of the art.