Deep neural network based robust computer-aided cataract diagnosis system using fundus retinal images

Highlights

• A deep neural network based robust cataract diagnosis system is proposed.

• The proposed cataract diagnosis system is robust against noisy environment.

• A combined feature extraction technique is used to enhance the robust performance.

• Noise level estimation-based classifier is presented for optimum performance.

• The proposed system is robust and better when compared with existing systems.

Abstract

Cataract is the cloudiness present in the eye lens due to denaturation of active protein cells. Cataract affects the quality-of-life and thereby troubling the daily routine activities. Early diagnosis and treatment may reduce the vision loss and delays the cataract progression. To diagnose large-screen population, the computer-aided cataract diagnosis (CACD) system using fundus retinal (FR) images is required. In this paper, a CACD system using FR images is proposed to achieve better diagnostic accuracy. It is perceived that the performance of existing CACD systems is poor against noisy input FR images. However, the distortion such as noise is unavoidable in input images due to complex processes involved in the image acquisition. Hence, it is required to consider the effect of noise in the design of CACD systems. So the proposed CACD system includes this issue in the design and provides the robust performance. In the presented CACD system, the features are extracted using combined feature extraction (CFE) technique using two independently fine-tuned deep convolutional neural networks. The noise level estimation (NLE)-based classification is adopted in the classification stage. In NLE-based classification, a set of multi-class support vector machine (SVM) classifiers, which are trained independently at noise levels from 0 to 25 are considered. Finally, the features extracted using CFE are then mapped to a specific multi-class SVM classifier based on noise level present in an input FR image. From the experimental results, it is observed that the proposed system exhibits superior performance than existing CACD systems under noisy conditions.

Introduction

Visual impairment is the inability of a human to perceive a clear vision. Out of 7.3 billion global population, 2.2 billion population suffering from visual impairments in which one billion visual impairment could have been prevented [1]. The addressable cataract cases out of one billion visual impairment population are approximately 65.2 million [2]. So cataract is a global problem. The prevalence of visual impairment is high in developing countries than in developed countries [3]. In the year 2015, the vision loss expert group (VLEG) estimated the global blindness and moderate to severe visual impairment (MSVI) statistics worldwide [2], [3]. As per VLEG, the estimated global blindness (MSVI) is 36 (217) million. Out of 36 (217) million of global blindness, worldwide 12.6 (52.6) million of blindness (MSVI) is due to cataract. By 2020, the blindness (MSVI) due to cataract is predicted to increase to 13.4 (57.1) million [2]. The uncorrected refractive error (URE), cataract, diabetic retinopathy (DR), age-related macular degeneration (AMD), glaucoma, etc. are considered as global causes of visual impairment [2]. The percentage of global causes for visual impairments such as MSVI and blindness are shown in Fig. 1. From Fig. 1, it is clear that the cataract is the leading cause of blindness and the second leading cause of MSVI.

The eye lens is transparent and responsible for clear vision. The optical homogeneity of an eye lens is due to several interdependent factors. Cataract is the haziness due to denaturation of active proteins in the eye lens. The opacity present in the eye lens scatters the light entering into the retina [4]. The common risk factors of cataract are age, ultraviolet-B exposure, diabetes, and genetic composition. But in most of the cases, older adults (age $>$ 50) are affecting due to cataract [5]. Cataracts can impact visual quality-of-life by affecting daily activities of human beings related to mobility, perceptibility, and societal life [6]. The treatment for cataract is surgery in which the cloudy lens is replaced with an artificial intraocular lens. The availability of eye care services are not equally available to all regions [7]. The unequal access to eye care leads to an increase in the waiting list such that surgeries performed are not meeting the demand at a given time. So cataract cases are increasing for every year [8]. The main challenge is to prevent or delay cataract formation. The early diagnosis of cataract can control vision loss worldwide.

In the process of diagnosis, the ophthalmologist needs to observe retinal image in order to provide grading for the cataract. The grading is the standard procedure in which the severity of cataract can be identified using standard grading methods. The commonly used grading method is the lens opacity classification system-III [9]. The traditional methods of cataract diagnosis are bulky, time-consuming and often requires a well-trained operator [10]. In recent years, the use of fundus cameras for cataract diagnosis is increasing due to various reasons such as flexibility, easy-to-use, low-cost, and operability with less-trained staff [10]. The fundus retinal (FR) image-based computer-aided cataract diagnosis (CACD) systems may enable the telemedicine application remotely and thus make it possible large-scale screening population [11]. Motivated by the preceding discussion, an effective CACD system using FR images is proposed in this paper.

Noise in FR images is inevitable due to complex image acquisition system [12]. It is observed that the performance of existing CACD systems [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30] is poor due to noisy input FR images. Hence, the proposed CACD system is intended to achieve better performance for good-quality and noisy input FR images. The presented CACD system consists of two stages: feature extraction and classification. In feature extraction stage, the combined feature extraction (CFE) technique is used. The CFE technique extracts features from two independently fine-tuned deep convolutional neural networks (DCNN) and then combines those features for better discrimination. The noise level estimation (NLE)-based classification using multi-class support vector machine (SVM) classifiers is used in the classification stage. The NLE-based classification consists a set of independently trained multi-class SVM classifiers at noise levels ranging from 0 to 25 in which a single classifier is selected based on NLE from an input FR image. To analyze the presented CACD system effectively, the FR images are collected from the publicity available EyePACS dataset [31]. The collected FR images are refined for good-quality and then divided into four categories such as normal (non-cataractous), mild, moderate and severe with the help of experienced ophthalmologists. The noisy FR image datasets are created using good-quality FR image dataset and then used in the analysis. The noise models such as additive white Gaussian (signal-independent), multiplicative and practical (Poisson-Gaussian) noise models are used in this work. The proposed CACD system is constructed using a signal-independent noise model and then made applicable to signal-dependent noises such as multiplicative and Poisson-Gaussian. The performance of the proposed CACD system is observed as better in terms of diagnostic accuracy when compared with existing CACD systems under noisy conditions.

The remainder of this paper is structured as follows: Section 2 describes the existing systems of CACD. The proposed CFE technique and NLE-based classification are explained in Section 3. The experimental setup required to analyze the proposed CACD system is presented in Section 4. The merits and demerits of the proposed CACD system are discussed in Section 5. The conclusion is presented in Section 6.