| Diabetic retinopathy is the most common complication of diabetes,and it is also one of the main diseases that cause blindness in humans.Because this is an irreversible eye disease,patients need to have timely and regular retinal examinations to avoid deterioration of the disease and eventually blindness.However,as the number of patients with diabetic retinopathy continues to increase,the current diagnostic model that completely relies on the ophthalmologist to manually segment the patient’s fundus biomarkers and then classify the disease has become unable to cope with it,especially when medical resources are scarce and medical Areas with insufficient levels and serious imbalances in the ratio of doctors to patients.Therefore,this paper designs and implements an auxiliary diagnosis system for automatic segmentation of biomarkers from fundus images,which is used to improve the diagnosis efficiency and quality of ophthalmologists.In color fundus images,hard exudate is a typical sign of a person suffering from diabetic retinopathy,and the key to judging whether the stage of retinopathy is proliferative or non-proliferative is whether there are new blood vessels in the retina.Therefore,this paper proposes an automatic segmentation model for blood vessels and hard exudates in fundus images based on the deep learning method,and combines Web technology to design a medical auxiliary diagnosis application system for ophthalmologists.The specific research contributions are as follows:(1)Aiming at the insufficient ability of the existing automatic retinal vessel segmentation model to segment small blood vessels and diseased areas in color fundus images,we propose an automatic retinal vessel segmentation method based on an improved holistically nested edge detection network(HED)network.Combining the super feature extraction ability of cavity convolution and residual deformable convolution and the feature fusion ability of HED network greatly improves the sensitivity and anti-interference of blood vessel segmentation in fundus images.(2)Aiming at the problem that the overall performance of the existing automatic segmentation models of hard exudates is not high,an improved U-shaped network model based on the attention mechanism is proposed for automatic segmentation of hard exudates in fundus images.The accuracy of the segmentation model is improved by adopting the dual attention mechanism of position and channel,and the technology of transfer learning is introduced to further improve the sensitivity of model segmentation.(3)Based on the automatic segmentation model of blood vessels and hard exudates in color fundus images designed by the first two research institutes,combined with Web technology,an auxiliary diagnosis system for automatic segmentation of fundus images of patients with diabetic retinopathy has been realized.By using the auxiliary diagnosis system,not only can the ophthalmologist’s diagnosis process be greatly accelerated,but also the diagnosis accuracy can be improved to a large extent. |
