Retinal Image Analysis using Image Processing Techniques /

Abstract

Retinal images provide a simple non-invasive method for the detection of several silent ocular diseases such as diabetic retinopathy and age-related macular degradation. Silent ocular diseases have early symptoms that slowly advance, and are usually unobserved by the patients. In the late stages, these diseases become recognizable after severe vision impairments have already occurred within the patient’s retina. Early detection and treatment of the silent ocular diseases can reduce their progression and improve the quality of life for their patients. However, many factors can result in the degradation of the images’ quality, such as the retinas’ structure, pupil dilation, ocular media opacity, and the experience of the operating person. In case of poor quality retinal images, such early disease symptoms, could be hardly detectable by the computer aided diagnosis (CAD) systems or the ophthalmologists leading to false negatives or misdiagnosis of the silent ocular diseases. Accordingly, retinal image enhancement is an essential preprocessing step necessary to increase the reliability of the performed diagnosis by improving the overall appearance of the different retinal structures within the retinal images. Wavelet transform has the benefit of being consistent with the human visual system in identifying the retinal structures. In addition, wavelet transform is a multi-resolution technique that reveals the finer image details related to the different retinal structures within its subsequent levels. It separates the retinal images’ luminance and edge information in its approximation and detail subbands, respectively. Thus, wavelet decomposition allows for the improvement of the contrast and illumination through manipulation of its low frequency approximation subbands. In addition, it permits for the enhancement of the retinal structures edges for sharper image as well as noise removal using its high frequency detail subbands. In this work, a wavelet-based retinal image enhancement algorithm is proposed that addresses the four most common quality issues within retinal images (1) contrast enhancement, (2) illumination enhancement, (3) noise removal, and (4) sharpness enhancement. Contrast and illumination enhancement involve applying contrast limited adaptive histogram equalization (CLAHE) and the proposed luminance boosting method to the approximation subband, respectively. Noise removal and sharpness enhancement are performed by processing the wavelet detail subbands, such that the upper detail coefficients are eliminated, whereas bilinear mapping is used to enhance the lower detail coefficients based on their relevance. Several analyses were performed for each of the four considered quality issues in order to tune their parameters. The proposed wavelet-based retinal image enhancement algorithm was tested on the public High-Resolution Fundus (HRF) dataset, which is characterized by its blurry, low contrast and poorly illuminated retinal images. Six different retinal image quality measures were considered to assess the proposed algorithm and to compare its performance against four other methods from literature. The comparison showed that the introduced method resulted in the highest overall image improvement followed by spatial CLAHE for all the considered quality measures. Thus, proving the superiority of the proposed wavelet-based enhancement method.