Research On Image Segmentation Of Diabetes Retinopathy Based On Deep Learning

Diabetes retinopathy is one of the most common complications of diabetes,and also one of the major eye diseases that cause blindness.Due to the increase in the number of patients in recent years,the workload of doctors has increased,and the heavy task of reading films has led to doctors being unable to provide timely feedback to patients.The focus accounts for a relatively small proportion of fundus images,which is extremely prone to misdiagnosis and missed diagnosis.Segmentation of diabetes retinopathy by computer algorithm can greatly reduce the workload of doctors,improve the efficiency of segmentation,and avoid the interference of human and subjective factors,which is of great significance for auxiliary diagnosis.Based on the multi-task learning method in deep learning,dissertation studies two different backbone networks and designs experiments for different types of lesions.The main work of this paper includes the following:Aiming at the problem of excessive resolution of fundus images and the relatively small proportion of lesions in fundus images.Most non lesion information pixels are removed using the Otsu threshold algorithm,and then the image is segmented into several small size images using a sliding window cutting method.Aiming at the segmentation problem of diabetes retinopathy(DR),a multi classification image segmentation method based on multi task learning is designed.First,the sub images without lesions are deleted to increase the proportion of sub images with lesions;Then,using the multi task learning attribute of UNet++,and using transposed convolution instead of traditional sampling,we perform multi output and multi focus image segmentation.Finally,we validate it on internationally published IDRi D and DDR datasets.Fundus microaneurysm is a key focus in the early stage of diabetes retinopathy.Aiming at the problem of segmentation of fundus microaneurysms,a method combining multi task learning network and multiple image pre-processing is designed.Firstly,by combining multiple image processing methods,the characteristics of microaneurysms are made more obvious;Secondly,the main task is to segment the image of microaneurysms,and the presence detection of microaneurysms is to be used as a secondary task.Multitasking combined with U-Net network is used to improve the segmentation effect of the main task;Finally,it was verified on the internationally open dataset IDRi D.The above two experimental results show that applying multi-task learning to diabetes retinopathy can improve the segmentation effect.