Three-stages hard exudates segmentation in retinal images

This paper proposes a three-stages method of hard exudate segmentation in retinal images. The first stage is the pre-processing. The color transfer is applied to make all retinal images to have the same color characteristics, based on statistical analysis. Then, only a yellow channel of each image is used in the further analysis. The second stage is the blob initialization. The blob detection based on color, size, and shape including circularity and convexity is used to identify initial pixels of hard exudates. The detected blobs must not be inside the optic disk. The third stage is the segmentation. The graph cut is iteratively applied on partitions of the image. The fine-tune segmentation in sub-images is necessary because the portion of hard exudates is significantly less than the portion of non-hard exudates. The proposed method is evaluated using the two well-known datasets, namely e_ophtha and DIARETDB1, in both aspects of pixel-level and image-level. Based on the comprehensive comparisons with the existing works, the proposed method is shown to be very promising. In the image-level, it achieves 96% sensitivity and 94% specificity for the e_ophtha dataset, and 96% sensitivity and 98% specificity for the DIARETDB1 dataset.