A random forest based approach for one class classification in medical imaging

Chesner Désir (1, 2) , Simon Bernard (2) , Caroline Petitjean (1, 2) , Laurent Heutte (2)

1 QuantIF-LITIS - Equipe Quantification en Imagerie Fonctionnelle
2 DocApp - LITIS - Equipe Apprentissage

In this paper, we address the problem of one-class classification for medical image classification. Indeed, in some situations, pathological samples may be difficult to acquire. In this case, one class classification (OCC) is a natural learning paradigm to be used. It consists in learning from only one class of objects, while two or more classes may be presented in prediction. We propose an original OCC method called One-Class Random Forest (OCRF), that combines ensemble learning principles from traditional Random Forest algorithm with an original outlier generation method. These two key processes complement each other for responding to OCC issues, and are shown to perform well on medical datasets in comparison to few other state-of-the-art OCC methods.

Mots clés: One-class classification; decision trees; ensemble methods; medical image classification; random forests

Domaines: Apprentissage [cs.LG]; Imagerie médicale

Chesner Désir, Simon Bernard, Caroline Petitjean, Laurent Heutte. A random forest based approach for one class classification in medical imaging. 3rd MICCAI International Workshop on Machine Learning in Medical Imaging (MLMI), 2012, Nice, France. pp.250-257. ⟨hal-00794160⟩