Hierarchical retinal blood vessel segmentation based on feature and ensemble learning

Highlights

• A supervised method based on feature and ensemble learning is proposed.

• The whole pipeline of the proposed method is automatic and trainable.

• Convolutional Neural Network performs as a trainable hierarchical feature extractor.

• Ensemble Random Forests work as a trainable classifier.

• Compared with state-of-the-arts, the experimental results are promising.

Abstract

Segmentation of retinal blood vessels is of substantial clinical importance for diagnoses of many diseases, such as diabetic retinopathy, hypertension and cardiovascular diseases. In this paper, the supervised method is presented to tackle the problem of retinal blood vessel segmentation, which combines two superior classifiers: Convolutional Neural Network (CNN) and Random Forest (RF). In this method, CNN performs as a trainable hierarchical feature extractor and ensemble RFs work as a trainable classifier. By integrating the merits of feature learning and traditional classifier, the proposed method is able to automatically learn features from the raw images and predict the patterns. Extensive experiments have been conducted on two public retinal images databases (DRIVE and STARE), and comparisons with other major studies on the same database demonstrate the promising performance and effectiveness of the proposed method.

Introduction

Retinal blood vessel segmentation has been widely used in various scenarios. For example, change of the retinal blood vessel appearance is an important indicator for various ophthalmologic and cardiovascular diseases, such as diabetes, hypertension, and arteriosclerosis [1], therefore, automatic segmentation and analysis of the retinal vasculature play an extremely vital role in the implementation of screening programs for diabetic retinopathy, the evaluation of retinopathy of prematurity, foveal avascular region detection, arteriolar narrowing detection, the diagnosis of cardiovascular diseases and hypertension, and computer-assisted laser surgery [2]. Moreover, the generation of retinal maps and detection of branch points have been utilized for temporal or multimodal image registration, retinal image mosaic synthesis, optic disc identification, fovea localization and biometric identification [2].

Both manual delineation and automatic algorithms have been used in retinal vessel segmentation. However, they have not gained wide acceptance due to several challenges. Manual delineation is skill demanding, tedious, time-consuming, and infeasible if given a large volume of fundus image databases. Accuracy of the automatic segmentation algorithms (to be reviewed in Section 2) is limited due to low blood vessel contrast, irregular shaped bright and dark lesions (in the form of hemorrhages, exudates, drusen and the optic disc boundary), intricate vessel topology (including vessel crossing and branching, as well as variation of vessel diameter and vessel grey levels) and nonuniform illumination of images as well as image deformation of scaling, skewing and other distortions.

In this paper, we present the hybrid method based upon convolutional neural network (CNN) [3] and ensemble random forests (RFs) [4] for automatic retinal blood vessel segmentation. We first employed a set of preprocessing steps to correct the nonuniform illumination of retinal images and to improve vessel contrast. We then used CNN to extract a set of hierarchical features which are not only invariant to image translation, scaling, skewing and other distortions, but also contain image based multi-scale information of the geometric structure of retina. We finally trained ensemble RFs to obtain a vessel classifier. The whole pipeline of the proposed method is trainable and automatic. Moreover, our method can effectively deal with the challenges of retinal vessel segmentation, as shown by our evaluations conducted using two publicly available databases (the DRIVE [5] and STARE [1]) and comparisons with state-of-the-art. Experimental results show that our approach is competitive with state-of-the-art by achieving sensitivity/specificity/accuracy/AUC values of 0.8173/0.9733/0.9767/0.9475 for the DRIVE database and of 0.8104/0.9791/0.9813/0.9751 for the STARE database. In contrast, previous methods of Fraz [6] produced values of 0.7406/0.9807/0.9480/0.9747 for the DRIVE database and of 0.7548/0.9763/0.9534/0.9768 for the STARE database.

The rest of this paper is organized as follows. Section 2 gives a review of related work reported in the latest literature. Section 3 provides the description of CNN and RF. Section 4 gives a detailed description of the proposed method. Section 5 evaluates the proposed method. Further, the concluding remarks are included in Section 6.