[Oral Communication] Méthode de Sélection des Caractéristiques basée sur le Random Forest : Prédiction de l’Evolution du Cancer de l’OEsophage en Imagerie TEP

Desbordes P^1,2, Ruan S¹, Modzelewski R^1,3, Vauclin S², Vera P^1,3, Gardin I^1,3
1 LITIS – QUANTIF – EA4108, Université de Rouen, France
2 DOSISOFT, Cachan, France
3 CENTRE HENRI BECQUEREL, Rouen, France

Presented at RFIA 2016

ABSTRACT
In oncology, many features can be extracted from PET images to describe tumours. We propose a new feature selection strategy based on random forests (RF) to detect the best predictive subsets of features. Our method is performed in 3 steps. First, a Spearman correlation analysis is carried out to eliminate the correlated features. Then, the achievement of importance rankings of the remaining features according to the RF is done. Finally, the best subsets of features are defined through a genetic algorithm (GA). Our method is evaluated by RF classification based on data from 65 patients with an oesophageal cancer. A comparison with other features selection strategies is performed. Our method gives more accurate results than those evaluated.