The Effect of Imbalanced Classes On Retinal Blood Vessel Segmentation

This study aims to segment the retinal blood vessels from the fundus image by considering the imbalanced classes of the training dataset. In the process of making the model, the majority class is undersampled by using random sampling while the minor class is oversampled by using SMOTE. The feature texture in the model is obtained from feature extraction using Gabor Filter. The experimental result shows that when the class is balanced, the sensitivity value increases compared to when the class is unbalanced. This study uses a model generated from Random Forest, Adaboost, and Support Vector Machine. The best segmentation result is obtained by using Random Forest which the accuracy of the method is 0.952, the precision is 0.729, the sensitivity is 0.718, and the specificity is 0.975.