SVM learning from imbalanced microanuerysm candidate datasets used feature selection by gini index

In the view of the characteristic of the imbalanced microanuerysm candidate datasets: a large number of negative samples, the different distributions of different classes and the irrelevant features exacted from each candidate for learning task, this paper proposes a feature selection algorithm that we selected the top features out of all features that were ranked in the increasing order of feature weights generated by Gini index, and then a modified SVM classifier is used to divide the microanuerysm candidates into two groups: true microaneurysms and false microaneurysms. The experiment on the training set of a publicly available database shows that the proposed new method has the best performance including the best free-response receiver operating characteristic (FROC) curve. Furthermore the proposed method based on top features selected by feature Gini index outperforms over all features.