Improving Supervised Microaneurysm Segmentation using Autoencoder-Regularized Neural Network

Abstract

This paper proposes the novel microaneurysm segmentation technique, based on the autoencoder-regularized neural network model. The proposed method is developed using two levels of the segmentation. First, the coarse-level segmentation stage locates the candidate areas using the multi-scale correlation filter and region growing. Second, the fine-level segmentation stage uses the neural network to obtain confidence values of candidate areas of being microaneurysm. The neural network based technique introduced in this paper is the modified multilayer neural network with an additional branch to take into account of the reconstruction error (in a similar fashion to the autoencoder). This modification to the neural network results in the consistent improvement in the classification performance, when compared to the conventional network without such modification. The proposed method is evaluated using the retinopathic online challenge dataset. It can deliver very promising results, when compared with the existing state-of-the-art techniques.