A New Approach for Detection of Retinal Haemorrhages in Colour Fundus Images

Abstract

We propose a new approach for retinal haemorrhages detection in a retinal image. Our approach is divided into three steps. Pre-processing step which consists of green and V band extraction, histogram matching, contrast enhancement and morphological opening operation is needed to improve the image quality. In the most important step, retinal haemorrhage which is a benchmark of diabetic retinopathy is extracted from colour retinal images using a two-dimensional matched filter. Finally, a post-processing step is performed to reduce the false positive of detected haemorrhages. Our approach is validated on 89 public retinal images from DIARETDB1 database. Three validation parameters, namely sensitivity, specificity and accuracy are calculated by comparing our result to its corresponding hand labelled haemorrhages image. The obtained average sensitivity, specificity and accuracy are 0.91, 0.98 and 0.98, respectively. This achievement is much better than that of other published methods.