Detection And Recognition Of Microaneurysms In Fundus Image Based On Deep Learning

With the development of medical equipment,people pay more attention to the fundus images.The effective analysis to the diagnosis can prevent the ocular fundus diseases.Diabetic retinopathy is a disease that caused by diabetes,and in the early stages of the disease the patients will appear microaneurysm in fundus images.By detecting the microaneurysm in fundus images,we can effectively promote the implementation of promoting the early screening of diabetic retinopathy.The purpose of this thesis is to analysis the detection and identification of microaneurysm in diabetic retinopathy,including image preprocessing,microaneurysm candidate detection,segmentation of retinal vascular image and microaneurysm identify,and then based on the cloud architecture I developed a microaneurysm detection diagnosis system.In this thesis the main work is as follows:(1)Fundus image preprocessing.Based on the analysis of fundus structure and fundus image,we extract the green channel,ROI region extraction,anisotropic diffusion filtering and CLAHE enhancement process to preprocess the fundus image.In this we can lay the foundation for image processing of later algorithms.(2)Microaneurysm candidate detection.By research on the morphological characteristics of the microaneurysm,I use template matching method to detect the microaneurysm candidate.Through analysis the defects caused by classic template matching method,this thesis proposes the DMPTM algorithm,namely based the classical template matching method we introduced template weights,radius and offset parameters to adjust the template matching process,the experimental results achieved 97.62% sensitivity in ROC fundus library,it improve the classical method.(3)Retinal vessel segmentation.This thesis studies and implements the fundus vascular morphological segmentation algorithm and segmentation algorithm of eye fundus blood vessel based on phase congruency,and then study the deep learning framework and U-net model,then apply the U-net model to the retinal vessel segmentation.In DRIVE fundus library,the experimental results is that under the curve of ROC the AUC value was 97.62%,the accuracy rate was 95.47%.(4)Identification of microaneurysm candidate.This thesis studies the popular framework of stream object detection and target recognition,including the RCNN model,Fast-RCNN model and the Faster-RCNN models,and ultimately use the Faster-RCNN model to detect microaneurysm,in ROC Library the experimental results achieved a sensitivity of 95.72%,specificity of 92.34%,accuracy of 93.67%.Finally we developed a retinal microaneurysm detection and recognition diagnosis system,and the system was tested to verify the feasibility and stability of the system.