AUTOMATIC DETECTION OF RETINAL HEMORRHAGE BASED ON GABOR WAVELET AND HYBRID KNNSVM ALGORITHM FOR FUNDUS IMAGES

Abstract

Retinal haemorrhage is the abnormal bleeding of the blood vessels in the retina, the membrane in the back of the eye. In retinal image, automated detection of haemorrhage is a major challenging factor. For automated detection of haemorrhage, a generalized framework is needed to train classifiers with optimal features learned from available dataset. Because of the variability in appearance of these lesions (i.e., haemorrhages), different techniques had been designed to detect each type of these lesions (i.e., haemorrhages) separately in detection system. We need a generalized framework to detect these types of lesions in fundus (i.e., retinal) image. A robust and computationally efficient approach for haemorrhage detection in a fundus retinal image is presented in this paper. Splat feature classification method is proposed with application to retinal haemorrhage detection in fundus images. Automated screening system is very much important to detect a retinal haemorrhages. Based on the supervised approach, fundus images are partitioned into non-overlapping segments covering the entire image. Each splat contains a similar colour and spatial location. A set of features is extracted from each splat using the GLCM & Gabor Wavelet. These features describe a characteristic relative to each pixel in a splat. Supervised classification predicts the likelihood of splats being haemorrhages with the optimal features subset selected in a two-step feature selection process. Preliminary feature selection is done by filter approach followed by a wrapper approach. Hybrid KNNSVM classifier is trained with expert annotation. From the resulting haemorrhages map, a haemorrhage index is assigned. A classifier could evaluate on the publically available dataset. This work will provide a greater AUC in splat level and image level. Our approaches can potential to be applied to other detection tasks.