A Novel Multi-Objective Based Feature Selection Method for Response Prediction

Abstract

Accurate response prediction is essential towards personalized treatment in radiation therapy. Excessive imaging features, extracted from medical images, pose a great challenge in radiomic analyses. Feature selection is an essential step to remove redundant and irrelevant features for model construction. We proposed a novel multi-objective based radiomic feature selection method (MRMOPSO), where the number of features, sensitivity, and specificity are jointly considered as optimization objectives for feature selection. The MRMOPSO innovated by three aspects: 1) Fisher score initialize the feature population to speed up the convergence; 2) Min-redundancy particle generation operations to reduce the redundancy between radiomic features, a truncation strategy was also introduced; 3) Particle selection operation guided by elitism strategies to improve local search ability of the algorithm. We evaluated the effectiveness of the proposed MRMOPSO method using a cohort of oropharyngeal cancer patients from The Cancer Imaging Archive (TCIA). 357 patients were used for model training and additional 64 patients were used for independent evaluation. The proposed methods were compared with (a) classical feature selection methods, i.e., Lasso, minimal-redundancy-maximal-relevance criterion (mRMR), F-score, and mutual information (MI), (b) single-objective feature selection methods, i.e., genetic algorithm (GA), particle swarm optimization algorithm (PSO) and (c) multi-objective feature selection methods, i.e., multiple objective particle swarm optimization (MOPSO), nondominated sorting genetic algorithm II (NSGA II). The other feature selection methods yielded AUCs, sensitivity, specificity of (0.48-0.71), (0.49-0.86), (0.33-0.67), respectively. The MRMOPSO achieved significantly highly AUC of 0.84 with smaller number of selected features on the independent dataset (Table 1). Additionally, the MRMOPSO remarkably improved the sensitivity (0.81), specificity (0.81) and achieved an excellent balance between sensitively and specificity. We demonstrated a novel multi-objective based radiomic feature selection method. The proposed algorithm effectively reduced feature dimension, and achieved superior AUC with simultaneous improved sensitivity and specificity, for radiomic response prediction.