Automatic method to distinguish manifestation areas of early diabetic retinopathy from image artefacts by using L*u*v* colour space
Abstract
Purpose
The RGB colour space was converted into seven different colour spaces: XYZ, CMY, HSL, HSV, HSI, L*a*b* and L*u*v*. The L*u*v* colour space presented optimal results, with the highest sensitivity and best reproducibility. We employed three-dimensional analysis of L*u*v* colour spaces to detect early diabetic retinopathy.
Methods
Six patients with small haemorrhages, hard exudates and photocoagulation marks were evaluated using fundus photography, which revealed image artefacts in the fundi of some patients. We constructed an experimental device similar to the optical system of a fundus camera, and created artificial eyes of the fundus, which were painted with four different colours. The image artefacts were photographed under each artificial eye using the experimental device. We analysed all images using Scilab 5.4.0 and SIVP 0.5.3 software. The software interpreted the values of the L*u*v* colour space as a three-dimensional graph, which was modified using a Gaussian filter.
Results
We calculated the difference between the manifestation and perimanifestation areas and image artefact and periartefact areas using the L*u*v* values. The L*u*v* values’ ratios of the image artefact to manifestation areas in the human eye were as follows: haemorrhage (6.2, 11.4, 7.4); hard exudate (3.2, 7.7, 2.5) and photocoagulation mark (8.1, 3.9, 6.2).
Conclusions
L*u*v* colour space is an effective mean of differentiating between small haemorrhages, hard exudates and photocoagulation marks from image artefacts.
