Automated Optic Disc region location from fundus images: Using local multi-level thresholding, best channel selection, and an Intensity Profile Model

Highlights

• Use of promising OD location reduces processing area in about 40%.

• Selects red or green channel with best range for Candidate generation.

• All images of the datasets are used for testing, given that no training is used.

• Column-wise intensity model reflecting natural variations of the optic disc zone simplifies classification.

• The method is image resolution independent by using relative measures.

Abstract

Background and objective

Location of optic disc, which corresponds to the visible part of the optic nerve in the eye, is of high importance for bright lesion detection of Diabetic Retinopathy by extracting it and avoiding false positives. Glaucoma detection processes details on the optic disc zone. Location of the macula uses optic disc location as a reference. Thus, the location of optic disc is relevant for several diagnosis procedures on retinal images. Several methods for OD detection in fundus images can be found in the literature; however, the issue is still open to reach better results in terms of accuracy, robustness and complexity. This work provides a simple and image resolution independent method for Optic Disc location for methods that use the optic disc zone elimination or extraction to perform some diagnosis.

Methods

This work proposes a simple and reliable method for OD region location in fundus images using four known publicity available datasets: DRIVE, DIARETDB1, DIARETDB0 and e-ophtha-EX. We are introducing an OD region location method based on OD’s characteristic high intensity and a novel method for feature’s extraction that aims to represent the essential elements that define an optic disc by proposing a model for the pixel intensity variations across the optic disc (column wise). The approach has four main stages: OD pixel region candidate generation, promising OD regions detection, promising candidate features extraction, and classification. All images from the four datasets were used for testing, since no training was used for classification.

Results

An OD location accuracy of 99.7% is obtained for the 341 retinal images within the four publicly datasets.

Conclusions

The obtained results show that the proposed method is robust and achieves the maximum detection rate in all four compared databases, which demonstrates its effectiveness and suitability to be integrated into a complete prescreening system for early diagnosis of retinal diseases. Use of promising OD region location reduces processing area in about 40%.

Introduction

Morphological detection of retinal structures such as the optic disc (OD), blood vessels, macula and fovea are a common step in most systems for automatic detection and screening of different retinal pathologies. Fundus images are used for diagnosis by trained clinicians to check for any abnormalities or changes in the retina. To alleviate physician work, images can be processed by an automated system that provides probable lesion areas, that the ophthalmologists will diagnose [1]. In particular, the detection of the OD, which corresponds to the visible part of the optic nerve in the eye, is an important task in retinal image analysis because it is a key reference for recognition algorithms [2], blood vessels segmentation [3,4], and diagnosing some diseases such as diabetic retinopathy (DR) [5,6] and for registering changes within the optic disc region due to diseases such as glaucoma [[7], [8], [9]] and the development of new blood vessels [10]. The OD is also a landmark for other retinal features, such as the distance between the OD and the fovea [11,12], which is often used for estimating the location of the macula [13] and is also used as a reference length for measuring distances in retinal images [14]. In addition, it is important to detect and isolate OD region because, most of the algorithms designed to segment/detect abnormalities such as hard exudates in DR will detect lots of false positives in OD region since the optic disc could present similar color, shape and size characteristics which can lead to potentially detect OD sections as exudates, negatively affecting the performance of the system [15]. Accurate identification of OD can be used to reduce the false positive rate while detecting the bright lesions [16].

Even though there is a significant variation in optic disc’s appearance and size, it’s mean diameter is clinically estimated on 1.5 mm [17] (near (1/30)^th the of retina’s area) and exhibits a distinctive appearance from the tissue zone that surrounds it due to the absence of pigment epithelium making OD’s color paler than its surroundings. In healthy fundus images, optic disc is usually seen as a bright circular or elliptic region crossed by a tree of vesins and arteries [18]. Although the OD main features and characteristics are relatively easy to describe, individual differences, diseases and other factors will influence characteristics of the optic disc. There are several factors that difficult the correct OD detection, for example, in some of the images the edges of the OD are not clearly visible, and some OD regions can be obscured due to the blood vessels that pass through it. Image quality can also affect the appearance of the OD. A retinal image may be unevenly illuminated or poorly focused, resulting in a less distinct and/or blurred OD. On the other hand, when there is presence of injuries such as exudates or severe problems in the lighting; the pixels that present the highest values of intensity in the image do not always correspond to the optic disc. Optic disc location methodology involves extensive research interest [19,20], due to its relevance for OD diseases and as an anatomic eye’s feature.

This work proposes a simple and reliable method for OD region location in fundus images. To make optic disc location robust, OD’s characteristic high intensity is combined with a model of the profile pixel intensity variation. The approach comprises four main stages: OD pixel region candidate generation, promising OD regions detection, promising candidate features extraction, and classification. The proposed method consists on a top-down approach, namely, we go from a group of high intensity local regions with coarse level features to finer features on extended regions with minimum requirements on size and intensity distribution. The main contribution of this work can be summarized in five aspects: (1) About 40% of fundus image reduction is achieved by defining a promising region to locate the OD. (2) Unlike most approaches, both channels red and green are considered. The channel who exhibits the best range for candidate generation is selected for further processing. (3) No training is required, so all the images of the datasets are used for testing and no additional datasets are needed to train a classifier. (4) A novel method for feature’s extraction is proposed, it aims to represent the essential elements that define an optic disc by proposing a model for the pixel intensity variations across the optic disc (column wise). (5) Since anatomical features calculated are relative to image’s original resolution, the results of the proposed approach were achieved without any alteration on the original size of the input images, ensuring that our detection algorithm is applicable to any image resolution available in fundus image datasets. The approach’s results proved to be reliable; especially for challenging images, including images with poor illumination, pathological changes, with dark OD due to uneven illumination and low contrast and images with bright exudates whose size and intensity are like OD. The proposed algorithm was tested in four publicly-available datasets achieving a competitive performance respect to other state-of-the-art methods.

The rest of the paper is organized as follows: in Section 2 the relevant and recent related work in detection of OD on fundus images is reviewed. In Section 3 the main stages of the proposed approach are described. In section 4 the characteristics of the public databases used are detailed. Section 5 shows the experimental results, the validation obtained using public databases, and a comparison with other methods from the literature. Finally, Section 6 provides the discussion and conclusions.