The Study On Glaucoma Diagnosis Method Based On Machine Learning

Glaucoma is a fundus disease with a high rate of blindness,which is an important public health problem affecting the field of visual health in China.Its concealment is very high and difficult to treat,so early diagnosis is essential.In order to supply more support and help in the accurate screening of glaucoma and reduce the burden of doctors,this thesis proposes an automatic auxiliary diagnosis system for glaucoma based on machine learning technology.The specific contents include:(1)Image preprocessing.Firstly,the mask of red channel in RGB space was extracted to carry out uniform size clipping.Then,the results of mean filtering,Gaussian filtering and median filtering were compared,and the median filtering was selected for contour smoothing and black edge removal.Finally,different image enhancement methods were compared in the L-channel of LAB space,and the contrast limited histogram equalization enhancement algorithm was selected to enhance the local contrast of the image.(2)Optic disc segmentation.Firstly,the time-consuming and accuracy of matched filtering,region growth and Otsu threshold segmentation methods were compared,and the Otsu threshold method was selected as the vascular segmentation method.Then,the main vessels were obtained by removing small connected domains and refining them.Then,the convolution optimization directional local contrast algorithm was used to extract bright spots,and the candidate area of optic disc was obtained.Finally,hough circle transform,active contour model,morphological operation and center point location were compared,and hough circle transform was selected as the optic disc segmentation method in this thesis.(3)Optic cup segmentation.Based on the located center point of optic disc,the optic cup was segmsed by super pixel segmentation,level set iteration and traditional threshold method.After a comprehensive comparison of the consumption time and the matching degree of the expert annotation effect,we finally choose the superpixel segmentation as the method of the eyeglass segmentation in this thesis.(4)Classification of glaucoma.Firstly,the cup-disk ratio was calculated and 15 texture features were extracted.Then,support vector machine,random forest,K-nearest neighbor algorithm and naive Bayes algorithm were used to test the classification of Drishti-GS1 and REFUGE fundus images.After comprehensive sensitivity,accuracy and other indicators,support vector machine was selected as the classifier of the scheme in this thesis.In order to verify the effectiveness of the proposed system,this thesis conducted a comprehensive analysis of Drishti-GS1 and REFUGE public databases from the aspects of classification accuracy,area under ROC curve and AUC value.The test results show that the proposed system can achieve high specificity of glaucoma diagnosis while ensuring sensitivity,and was superior to some existing algorithms in classification accuracy.