Creation of Retinal Mosaics for Diabetic Retinopathy Screening: A Comparative Study

Abstract

The creation of retinal mosaics from sets of fundus photographs can significantly reduce the time spent on the diabetic retinopathy (DR) screening, because through mosaic analysis the ophthalmologists can examine several portions of the eye at a single glance and, consequently, detect and grade DR more easily. Like most of the methods described in the literature, this methodology includes two main steps: image registration and image blending. In the registration step, relevant keypoints are detected on all images, the transformation matrices are estimated based on the correspondences between those keypoints and the images are reprojected into the same coordinate system. However, the main contributions of this work are in the blending step. In order to combine the overlapping images, a color compensation is applied to those images and a distance-based map of weights is computed for each one. The methodology is applied to two different datasets and the mosaics obtained for one of them are visually compared with the results of two state-of-the-art methods. The mosaics obtained with our method present good quality and they can be used for DR grading.

Detection and Mosaicing Techniques for Low-Quality Retinal Videos

“…This process aims to prepare and transform the data for a better overall result. Based on the [12] approach, this problem's correction was made by performing the normalization of all the dataset's image average intensity. The normalization of each image is described with Equation (1).…”

“…Due to different colors intensities or irregularities in the transition between images, the borders of the images meet will need to be smoothed. This step is known as image blending or stitching and can be divided into four classes: feathering-based, pyramid-based, gradientbase, and optimal seam-based blending, as presented in [12,13].…”

“…The next step is image blending. In this stage was implemented the featuring-based method as presented by Melo et al [12], which creates two weighted masks to achieve a smoother look to the transition between images. Firstly, it is found the intersection between the images and, for each pixel inside of the overlapping region, are applied the Equations ( 2) and (3).…”