Fundus Image Enhancement Using Hybrid Deep Learning Approaches

Imaging of the retinal fundus plays a main role in retinal disease diagnosis. Intricate information in retinal fundus images, such as small vessels, microaneurysms, and effusions, may appear unclear due to border artifacts, abrupt changes in colour levels, and uneven illumination which will result in poor diagnosis of retinal diseases. The existing retinal image enhancement methods mostly address either the vessel contrast or uneven illumination. Further preserving the edge features along with enhancing the contrast as well as adjusting the illumination is addressed in the proposed method. First, Haar Wavelet transform is applied to the retinal images followed by gamma adjustment to achieve a uniform and an improved illumination. Then the images are processed using a CofE-Net architecture with RSA and LQA modules to obtain an image along with basic background information. The image containing the background basic information is then merged with the original image to obtain a contrast improved fundus image. The proposed method achieves better results based on visual assessment as well quality parameters like PSNR, structure similarity measure, Enhancement measurement error and Edge based Contrast measurement.

I. Introduction

Enhanced retinal images clearly show detailed fundus information such as blood vessels and optic disc, for improving the performance of Computer Aided Diagnosis (CAD) systems such as Increased network for segmentation of retinal vessels[1].