Fundus Image Segmentation Using Step Time Signal Polynomial Fitting

Image segmentation is a basic and important strategy to extract the region of interesting (ROI). However the ROI of retinal fundus image, such as the foreground or optic disc, often could not be generated correctly, due to poor lighting environments, physiological features or improper imaging equipment. Hence, a novel fundus image segmentation method is proposed in this paper. Briefly speaking, considering each line of fundus image as a time series signal, we adopt a step time series signal polynomial curve fitting methodology to smooth the noise and to exclude all features in the fundus image without any loss of the target margin. After fitting the fundus image, a refinement tactic with Kmeans algorithm is used to obtain the final segmentation results. This unsupervised segmentation approach is a completely robust method, because all break points in the signal are detected automatically by the Buishand Range (BR) homogeneity breakpoint detection algorithm. To demonstrate the experimental performance of our new proposal, the test results of foreground segmentation have been compared with their ground truth on DRIVE and STARE datasets, and yield an encouraging performance. Based on these two public datasets and our private database which have been annotated by the medical experts, the experimental outputs of optic disk segmentation achieve the average overlap area ratio of 92.88%, outperforming existing unsupervised methods in this field.