Optic Disc Segmentation in Fundus Images Using Anatomical Atlases with Nonrigid Registration

Abstract

According to a WHO report, approximately 253 million people live with vision impairment, 36 million of which are blind and 217 million have moderate to severe vision impairment. In a recent estimate, the major causes of blindness are Cataract, Uncorrected refractive index, and Glaucoma. Thus in medical diagnosis, the retinal image analysis is a very vital task for the early detection of eye diseases such as Glaucoma, diabetic retinopathy (DR), Age-macular Degeneration (AMD) etc. Most of these eye diseases, if not diagnosed at an early stage might lead to permanent loss of vision. A critical element in the computer-aided diagnosis of Digital Fundus images is the automatic detection of the optic disc region. Especially for the Glaucoma case, where cup to disc diameter ratio (CDR) is the most important indicator for detection. In this paper, we present a nonrigid registration based robust optic disc segmentation method using image retrieval based optic disc model maps that detect optic disc boundaries and surpasses the state-of-the-art performances. The proposed method consists of three main stages: (1) a content-based image retrieval from the model maps of OD using Bhattacharyya shape similarity measure, (2) constructing the test image specific anatomical model using the SIFT-flow technique for deformable registration of training masks to the test image OD mask, and (3) extracting the optic disc boundaries using a thresholding approach and smoothen the image by applying morphological operations along with the final ellipse fitting. The proposed work has used three datasets RIM, DRIONS and DRISHTI with 835 images in total. Our average accuracy values for 685 test images is 95.8%. The other performance parameter values are Specificity is 95.54%, Sensitivity is 96.13%, Overlap is 86.46% and Dice metric is 0.924 respectively, which clearly demonstrates the robustness of our optic disc segmentation approach.