Fundus Image Vessel Segmentation And Exudate Detection Based On OD Localization

The physiological structure in the retinal fundus image can be directly observed by photographic instruments without trauma,including the normal blood vessels,optic discs and possible lesions,which has important research reference value for clinical medicine.The thickness of the blood vessels,the size and the change in brightness of the optic disc in the fundus image all reflect the changes in human function to some extent.Diabetic retinopathy,a retinopathy caused by diabetes,is one of the three major blind diseases in the world.Therefore,the identification of the physiological structure of the retinal fundus image,as well as the screening of potential lesion characteristics,was of great significance for reducing blindness caused by diabetic retinopathy.In this paper,combined with the research hotspot of retinal fundus image,the blood vessel segmentation,the optic disc localization,and the hard exudate detection of the fundus image are realized based on MATLAB.The main research contents are as follows:(1)A retinal vessel segmentation method based on semi-trained generalized linear model was studied and implemented.Firstly,the green channel image was preprocessed with CLAHE enhancement,mean filtering and multi-scale Gabor wavelet transform,and the results showed prominent vascular features and weak background noise.Then,based on the generalized linear model,Gabor features of half of the preprocessed image are extracted,and the model is trained and used for the segmentation of the whole image.Finally,combined with the grayscale maximum of the green channel for post-processing to reduce noise.The method was validated on the DRIVE database.The experimental results showed that the average accuracy was 0.9455,the sensitivity was 0.8724,and the specificity was 0.9337.This method was used to train half-images,which made up for the defects that traditional machine learning classifiers may not be applicable to all libraries.The experimental results showed that the sensitivity was much higher than other classification-based algorithms,and it achieved good segmentation effect in both normal and diseased fundus images.(2)A fast optic disc segmentation method based on background subtraction and Hough circle detection was studied and implemented.In the pre-processing stage,AHE image enhancement and morphological operation were used to process the red channel image;then the background subtraction was achieved by combining the red channel mean filtering results to obtain the shadow image of the optic disc.In the localization stage,the candidate regions of optic disc in the shadow image were screened by combining the size ratio of optic disc and vascular segmentation results,and the optic disc localization and segmentation were realized based on Hough circle detection.Combined with the HRF fundus library for verification,this method achieved an average overlap index of 0.87;combined with the self-defined rules for DRIVE database and diabetic retinopathy commonly data DIARETDBO and DIARETDB1 fundus library for optic disc segmentation,achieved]00%,96.15%and 95.5]%optic disc segmentation accuracy.This method combined the difference of optic disc and background brightness to realize the localization of optic disc,with fast segmentation speed and high accuracy,and also had good applicability in the image of diabetic retinopathy.(3)A hard exudate detection method based on CMY magenta channel dynamic threshold segmentation and support vector machine was studied and implemented.Firstly,the M channel image of the fundus image CMY color space was extracted as a pre-processed image,and the threshold segmentation was conducted on its histogram area.In addition,optic disc was removed based on the segmentation method of optic disc in this paper to obtain candidate regions for exudate detection.Finally,24-dimensional features of the exudate candidate region of the bottom image of DIARETDB1 library were extracted to construct SVM.In this paper,the sensitivity and prediction rate of 93.91%and 95.7%were obtained in the evaluation criteria based on the retinopathy area;in the image-based evaluation standard,the average sensitivity was 100%,the specificity was 89.3%,and the average accuracy was 94.6%.The method combined the color characteristics of the exudates for preprocessing,and adopted the traditional machine learning method to obtain a higher prediction rate and accuracy,avoiding the disadvantages that the deep learning method required a large amount of data support.