Exudates Segmentation In Retinal Images Based On Boosted Soft Segmentation Algorithm

Exudates is a primary sign of diabetic retinopathy (DR) which is a common retinal complication associated with diabetes and the most main cause of blindness. Hard exudates (HEs) have been found to be the most specific markers for the presence of retinal oedema, the major cause of visual loss in non-proliferative forms of DR and one of the most prevalent lesions during early stages of DR. Automatic segmentation of hard exudates (HEs) in fundus images is clinically significant for the prevention of vision loss with an early screening process.Fundus images permit a high quality for detecting early signs of DR and are widely used by ophthalmologists to study eye diseases like diabetic retinopathy. Automatic HEs segmentation in fundus images is a difficult task due to the uneven illumination, poor contrast, color variation of the retinal images and size variation. In addition, the optical disc (OD) is very similar to HEs.In this paper, we propose an effective framework to automatically segment hard exudates (HEs) in fundus images. Our framework is based on a coarse-to-fine strategy, as we first get a coarse result allowed of some negative samples, then eliminate the negative samples step by step. In our framework, we make the most of the multi-channel information by employing a boosted soft segmentation algorithm. Additionally, we develop a multi-scale background subtraction method to obtain the coarse segmentation result. After subtracting the optical disc (OD) region from the coarse result, the HEs are extracted by a SVM classifier. The main contributions of this paper are:(1) propose an efficient and robust framework for automatic HEs segmentation; (2) present a boosted soft segmentation algorithm to combine multi-channel information; (3) employ a double ring filter to segment and adjust the OD region. We perform our experiments on the pubic DIARETDB1 dateset, which consists of 89 fundus images. The performance of our algorithm is assessed on both lesion-based criterion and image-based criterion. Our experimental results show that the proposed algorithm is very effective.