Optic disk detection and segmentation for retinal images using saliency model based on clustering

Optic disk (OD) is considered one of the main features of a retinal image. OD detection plays an important role in retinopathy analysis. In this study, a new OD detection method using saliency model based on clustering is presented to simulate the human filtering mechanism of visual system for OD detection of fundus images. First, the candidates of OD regions are extracted from fundus images using k-means clustering. Second, two saliencies of sub-regions are computed, and the maximum saliency region from the image is selected as the OD region. Third, the original OD contour can be extracted by ellipse fitting after detecting the convex hull of the OD. Finally, the OD contour can be accurately segmented by active contour. A test is performed with 1422 colored fundus images from four different colored fundus image databases. Experimental results indicate that the detection accuracy for OD is up to 94%, and the segmentation accuracy is up to 88% for Drishti-GS database. The proposed method effectively overcomes the influence of large bright lesions on OD detection and is applicable to incomplete ODs. The method also does not rely on vessel segmentation, which results in short computation time. This study also confirms the effectiveness and robustness of the proposed algorithm. © 2018 Computer Society of the Republic of China. All rights reserved.