An Ensemble Retinal Vessel Segmentation Based on Supervised Learning in Fundus Images

An ensemble method based on supervised learning for segmenting the retinal vessels in color fundus images is proposed on the basis of previous work of Zhu et al. For each pixel, a 36 dimensional feature vector is extracted, including local features, morphological transformation with multi-scale and multi-orientation, and divergence of vector field which is firstly used to extract feature of retinal image pixels. Then the feature vector is used as input data set to train the weak classifiers by the Classification and regression tree (CART). Finally, an AdaBoost classifier is constructed by iteratively training for the retinal vessels segmentation. The experimental results on the public Digital retinal images for vessel extraction (DRIVE) database demonstrate that the proposed method is efficient and robust on the fundus images with lesions when compared with the other methods. Meanwhile, the proposed method also exhibits high robustness on a new Retinal images for screening (RIS) database. The average accuracy, sensitivity, and specificity of improved method are 0.9535, 0.8319 and 0.9607, respectively. It has potential applications for computer-aided diagnosis and disease screening.