Vessel identification in diabetic retinopathy.

Abstract

Diabetic retinopathy is the single largest cause of sight loss and blindness in 18 to 65 year olds. Screening programs for the estimated one to six per- cent of the diabetic population have been demonstrated to be cost and sight saving, howeverthere are insufficient screening resources. Automatic screen-ing systems may help solve this resource short fall. This thesis reports on research into an aspect of automatic grading of diabetic retinopathy; namely the identification of the retinal blood vessels in fundus photographs. It de-velops two vessels segmentation strategies and assess their accuracies. A literature review of retinal vascular segmentation found few results, and indicated a need for further development. The two methods for vessel segmentation were investigated in this thesis are based on mathematical morphology and neural networks. Both methodologies are verified on independently labeled data from two institutions and results are presented that characterisethe trade off betweenthe ability to identify vesseland non-vessels data. These results are based on thirty five images with their retinal vessels labeled. Of these images over half had significant pathology and or image acquisition artifacts. The morphological segmentation used ten images from one dataset for development. The remaining images of this dataset and the entire set of 20 images from the seconddataset were then used to prospectively verify generaliastion. For the neural approach, the imageswere pooled and 26 randomly chosenimageswere usedin training whilst 9 were reserved for prospective validation. Assuming equal importance, or cost, for vessel and non-vessel classifications, the following results were obtained; using mathematical morphology 84% correct classification of vascular and non-vascular pixels was obtained in the first dataset. This increased to 89% correct for the second dataset. Using the pooled data the neural approach achieved 88% correct identification accuracy. The spread of accuracies observed varied. It was highest in the small initial dataset with 16 and 10 percent standard deviation in vascular and non-vascular cases respectively. The lowest variability was observed in the neural classification, with a standard deviation of 5% for both accuracies. The less tangible outcomes of the research raises the issueof the selection and subsequent distribution of the patterns for neural network training. Unfortunately this indication would require further labeling of precisely those cases that were felt to be the most difficult. I.e. the small vessels and border conditions between pathology and the retina. The more concrete, evidence based conclusions,characterise both the neural and the morphological methods over a range of operating points. Many of these operating points are comparable to the few results presented in the literature. The advantage of the author's approach lies in the neural method's consistent as well as accurate vascular classification.