Deep-learning based automated segmentation of Diabetic Retinopathy symptoms

Purpose: Retinal fundus images are an important basis for reflecting retinal health status, are widely used in clinical diagnosis, and have important significance in fundus image processing and analysis. At present, clinicians rely on manual examination of fundus images when diagnosing retinopathy, which is both time- and labor-consuming. However, the need for rapid auxiliary diagnostic imaging is increasing day by day. Therefore, this study evaluates deep learning as a preprocessing method for optic disc examination along with the method proposed in this study. Methods: During optic disc segmentation, the optic disc region in original images is blurry and unclear compared to the entire image. The halo around the optic disc causes boundaries to become unclear and makes examination difficult. In this paper, different preprocessing methods were used to solve the problem of blurry optic disc region and the effectiveness of the various preprocessing methods was evaluated. Preprocessing methods used in this study to identify optic disc regions that were not apparent in the original image include the original image, green channel, CLAHE, and subtraction of average filter from Gaussian filter (sigma=1) using the original image. This preprocessing method is known as local differential filter and was uploaded to U-Net for training. Results: Evaluate the proposed method on the MICCAI REFUGE Challenge 2018 database. In the performance of Dice, the original image is 0.9473; the LDF image is 0.9521; the green channel image is 0.9429; CLAHE image is 0.9499. Conclusions: Currently, deep learning is used in many types of preprocessing for segmentation. In this study, we preprocessed fundus images and inputted them into the model for training. Finally, LDF image was used to obtain the best preprocessing method for optic disc segmentation in fundus images.