Research On Quality Evaluation And Segmentation Of Fundus Image For Assistant Medical Diagnosis

The fundus image is obtained by special fundus camera,including retinal vascular tissue,optic disc cup,macula and other major structures.It is a kind of medical image widely used in auxiliary medical diagnosis.In clinic,the position and size of the optic disc/cup tissue on the fundus image can assist the ophthalmologists in diagnosing glaucoma,observing the change of macula can predict the senile maculopathy in advance,while diabetic retinopathy can help to judge by observing whether there is hard exudate and other exudates on the fundus image,and also by observing the retinal vascular structure help doctors diagnose some systemic cardiovascular diseases.Therefore,in order to make the fundus image better assist the medical diagnosis,this paper focuses on the evaluation and segmentation of fundus image in the following three aspects:(1)In this paper,a method for evaluating the unreferenced quality of retinal image based on vascular segmentation is proposed.Firstly,a database is constructed,which includes two kinds of distortion types and eight levels of distortion degree,and the vascular segmentation accuracy corresponding to each distortion image is taken as the label of the image;secondly,the statistical characteristics of pixel value,image texture features and vascular shape features of the retina image are extracted to form the feature vector;finally,the feature vector is extracted according to the feature vector The final retinal image quality was obtained by regression method.(2)In this paper,we propose a combined segmentation method based on super-pixel segmentation.In the first part,the region of interest is extracted,that is,the rectangular frame containing the video disk region is cut out,and then the region of interest is divided into different super-pixel blocks by multi-scale super-pixel segmentation;in the second part,according to the characteristics of the color and spatial structure of the video disk,each super-pixel block is extracted,including features such as color channel and space,to form a feature vector,using random forest In the third part,on the basis of the visual disc segmentation,except for the visual disc region,the image is erased to reduce the influence of blood vessels,and the cup is roughly segmented by the method of fuzzy c-means clustering,and finally the cup is segmented by some post-processing technologies The final result is obtained by fine segmentation.(3)Based on the structured learning algorithm,a method is proposed to segment the vascular tissue,optic disc structure and hard exudate pathological tissue of retina automatically.Firstly,the region of interest is extracted,the regions of interest of blood vessel,optic disc and hard exudate structure are extracted respectively.The extraction of these regions will reduce a lot of interference for the subsequent fine segmentation of various structures.Secondly,according to the structural characteristics of blood vessel,optic disc and hard exudate,the features of these three types of structures are extracted,and then based on these features,the structural features are extracted Learning to train three kinds of structure to get the classifier model,based on the corresponding model to achieve the edge mapping of the structure;thirdly,according to different structure,using the corresponding post-processing method to get the final segmentation results.