Robust feature selection to predict lung tumor recurrence

Abstract

The recurrence of cancer increases the risk of death. The ability to predict such recurrence could be beneficial to planning treatment.We aim to find a predictive feature subset from a series of spatiotemporal PET image characteristics, including SUV-based and texture features, in order to predict lung tumor recurrence one year after treatment. To overcome the small sample size, class imbalance problem, we propose a hierarchical forward selection algorithm to select the smallest feature subset that results in the best prediction performance. As the SUV-based features have been recognized as significant predictive factors for a patient’s outcome, we propose incorporating this prior knowledge into the selection procedure to improve its robustness and accelerate its convergence. By fixing the first feature as one SUV parameter, the proposed hierarchical forward selection yields a small robust feature subset with promising prediction performance.