Detection of exudates from retinal images using morphological compact tree

Exudates are an important sign of Diabetic Retinopathy. Retinal Exudates are formed when lipid leakage occurs from damage capillaries. They are deep yellowish in colour and can be easily confused with other yellowish regions in retina. Detection of exudates is very important for developing an automated screening system for detection of diabetic retinopathy. In this paper we focus on detection of exudates through morphological compact tree. We did some pre-processing for removal of noise and enhancement of image. Blobbing technique was applied and all the connected pixels were counted as single blob. These blobs are then passed through an area filter which removes blobs with very large areas. The remaining blobs are then divided into three categories small, medium and large. The medium and large blobs are again fed into pre-processing mode one by one to remove strong boundaries effect and extract the exact suspected candidate location. All the blobs are then passed through morphological compact tree of filters which removes the non-exudates regions through different. For each image different set of threshold values for the filters are required. In our technique we are setting it manually but further research is needed to find out the optimal threshold values or a technique which can calculate adaptive thresholds values for these filters. This is very simple method for the detection of exudates as it uses only a morphological filtration tree. 10 images of dimension 500∗752 were analysed in this experiment. The results were compared with ophthalmologist's hand drawn ground truths. Mean recall of 78 percent and mean precision of 56 percent were obtained.