Detection Of Retinal Microaneurysm In Diabetic Retinopathy

Diabetic retinopathy(DR) is an eye complication caused by diabetes,if not prevented as soon as possible,there will be a higher risk of blindness.There-fore,early screening through fundus images becomes very important.With the increase of the number of patients and the sharp increase of fundus images,oph-thalmologists are tired of viewing the fundus images.Therefore,automatic fundus image lesion detection can effectively improve the quality and efficiency of screen-ing of images,and greatly reduce the cost of diagnosis and treatment of patients.Hence very valuable for both doctors and patients.This paper focuses on the early DR lesion of microaneurysm,applied a series of algorithms to detect and recognize mcroaneurysms.Firstly,this paper introduce the relationship between diabetes and diabetic retinopathy,as well as its etiology and clinical manifestations.And DR is di-vided into two classes and six stages,the early symptom is microaneurysm.This paper introduces the eye structure and fundus imaging method,highlighted the importance of the automatic detection method in fundus images.The research content and research methods are introduced,and the literature methods are summarized comprehensively.Listed datasets for DR lesions detection and seg-mentation,and its lesion types,corresponding image numbers.Also,introduce various measurement indexes,the following three aspects of innovative work have been done in the task of microaneurysms detection.(1)Based on Canny edge detection operator,a line interpolation algorithm is designed to segment FOV region.In the FOV region of the image,the mask is generated and the surrounding black area is removed by clipping.With failure of traditional hybrid thresholding segmentation method in local highlight or dark area,based on the human visual perception of the edge contour,the edge detec-tion line interpolation algorithm in this paper can effectively solve the problem of segmentation failure in local bright or dark area by hybrid threshold algorithm,The experimental results and the graph 3.4 show that the proposed algorithm is robust to images with edge gradient.(2)Proposed MA detection algorithm based on directional derivative weighting and local contour analysis to extract microaneurysm candidates.Different prepro-cessing methods are used for two candidate detection algorithms,The weighted image with Gaussian directional derivative convolution kernel is used for iterative threshold algorithm to obtain MAs,as well as the local minimum algorithm.The Gaussian gradient convolution kernel in weighted image is improved to be direc-tional derivative convolution kernel,which can effectively improve the contrast of image lesions.The improved block filtering algorithm eliminates a large number of Non-MAs,increases the screening conditions according to the characteristics of microaneurysms,and takes the union of the two extraction results as final can-didates.which results in a higher recall rate while greatly reducing the number of false positives of the block filtering algorithm.(3)Based on several models improved feature expressions.The norm is used to describe the difference between Gaussian fitting curve and the mean profile bsed on the local section line of microaneurysms,the rank and singular value of the profile matrix are added as features.The support region of candidate points is expanded by weighted band-pass convergence index,and the calculation region is expanded to the interior of the support band to investigate the comprehensive contribution of the dark area in the image to the convergence index Finally,the rusboost classifier is used to verify the classification performance on e-ophtha-MA and diaretdb1 datasets,and the AUC values are 0.9679 and 0.956 respectively.ROC and FROC curves show that the proposed algorithm has good classification ability,and obtained the highest fscore on the e-ophtha-MA dataset,also the highest sensitivity on some preset FPI values.Although Fscore is slightly lower than the optimal method on DIARETDB1 dataset,but this paper obtains the highest sensitivity of 0.77451 when FPI is 8.This paper proposed a mask generation algorithm based on Canny edge detec-tion,improved the candidate extraction algorithm based on iterative thresholding and local minimum,obtained a higher recall rate in candidate extraction with im-proved block filter,the false positive number is greatly reduced compared with other literatures.multiple features are improved based on the norm of local pro-files,profile matrix and local convergence index.Finally,RUSBoost iterative class balanced sampling classifier achieved the highest F-score in e-ophtha-MA dataset,which verified the preprocessing,candidate extraction and multiple fea-tures proposed in this paper have better microaneurysms detection ability.