- October 17, 2016
- Category: Nuclear Medicine, Scientific Publications
Desbordes P1,2, Modzelewski R1,3, Ruan S1, Pineau, P2, Vauclin S2, Vera P1,3, Gardin I1,3
1 Litis – QuantIF – EA4108, University of Rouen, France
2 Dosisoft, Cachan, France;
3 Henri Becquerel Center, Rouen, France
Presented at MICCAI Conference 2016
Prediction of cancer evolution is an important issue for adaptation and evaluation of a therapy. Many features can be extracted from PET images to describe cancer. However which ones are relevant for prediction? In this paper, we propose a feature selection strategy based on random forest to select features having a prognostic or predictive value among a large amount of dierent kinds of characteristics. Our method is performed in 3 steps. First, a Spearman rank correlation is carried out to keep uncorrelated features. Then, a random forest algorithm is applied to nd the most relevant subset of features. Our method is evaluated on a PET database of 66 patients with an oesophageal cancer using two classiers: support vector machine and random forest. Results show an improvement of the classication accuracy in a range of [4.2 to 16.5%] compared to when using all features without selection. These results are compared with a classical feature selection method (SVM-SFFS) showing that our method gives better results than SVM-SFFS.