Eye Lesion Detection Via Feature Modeling Of Retinal Soft Exudation

Diabetic retinopathy is one of the four major blinding diseases in the world.Diagnosis of it can prevent the disease and confirm the severity of the disease,so as to realize the diagnosis and treatment in line with the disease.However,traditional diagnostic methods are inefficient and require doctors to have sufficient clinical experience to make judgments.When computer-aided detection is used,physiological structures such as blood vessels,optic disc,and macula have similar color,texture and other information as diabetic retinopathy,and it is not easy to achieve lesion segmentation.In particular,the soft exudation produced in the diseased stage is sometimes easily confused with structures such as hard exudation and optic disc due to its own morphological characteristics and texture characteristics,and is sometimes difficult to find in the background of the fundus.Therefore,implementing a lesion detection model for soft exudates is a challenging and realistic work.In this thesis,the algorithm of lesion detection for soft fundus exudation is studied.The main work is as follows:（1）A segmentation and extraction method of soft exudation candidate regions is proposed.The field of view template in the fundus image is found by threshold segmentation,the fundus blood vessels are quickly extracted by Gaussian matched filter,and the optic disc is segmented by the structural characteristics of blood vessels and optic disc.Morphology realizes the extraction of candidate regions,and finally removes the region where the optic disc is located to obtain lesion candidate regions.（2）A retinopathy detection method based on multi-feature joint representation is proposed.The obtained candidate regions are respectively extracted with deep features and hand-crafted features through shallow convolutional networks and hand-designed operators.After dimensionality reduction and feature fusion,they are input into random forest classifiers for training,and finally a lesion detection model is obtained.In the DDR data set,the comprehensive evaluation index F-score of this algorithm increased from0.7073 to 0.9188;in the IDRID data set,the F-score index of this algorithm increased from 0.8702 to 0.9147.（3）A semantic information-based retinopathy detection method is proposed.Based on the idea of pixel-level segmentation through encoder-decoder structure in semantic segmentation,this thesis designs a lesion detection model through Atrous Spatial Pyramid Pooling（ASPP）,Xception network and attention mechanism.The model achieves an Fscore of 0.9320 on the DDR dataset,which is close to the best-performing 0.9339,and improves the F-score from 0.8982 to 0.9059 on the IDRID dataset.