Feature Modeling And Pathological Detection For Retinal Microaneurysm

Diabetic retinopathy（DR）is a common complication of diabetes.Patients will suffer from vision problem or even blindness if they do not pay any attention to it in the beginning.The detection of microaneurysm（MA）,which is the earliest symptom of DR,is critical for the prevention and early treatment.Currently,the diagnosis of MA is simply based on the professional advices from ophthalmologists.This method is obviously subjective and leads to lots of misdiagnosis.Computer-aided diagnosis is a good solution to handle this problem,which can not only improve the detection accuracy,but also effectively reduce the misdiagnosis.It is also benifitical for saving medical resources.The thesis uses the digital image processing technology to detect the MAs from two kinds of data sources,i.e.,the color fundus and fundus fluorescein angiography（FFA）images,and aids the doctors to gives the final diagnosis.The thesis is organized as follows:1.The theoretical basis and background of DR and MA detection are studied,including the public datasets,the theory and the method of feature fusion and feature dimensionality reduction,machine learning classifier model,evaluation of medical image processing field,and so on.2.The MA detection method of color fundus image based on saliency detection and multi-feature modeling is proposed.Firstly,the saliency detection algorithm via graphbased manifold ranking is used to obtain the detection masks.Secondly,combined with gray analysis,the region detection based on morphology and frequency-tuned region detection is used to extract and optimize the candidate regions.Then,the features of geometry,intensity,contrast,texture,multi-scale Gaussian filtering response and adaptive sliding band detector response are extracted.The sparse principal component analysis is used for feature fusion and feature dimensionality reduction,and finally support vector machine is used to obtain the real and reliable lesion regions.Based on E-ophtha dataset,the sensitivity,positive prediction value and F1-score of the algorithm are 0.9501,0.9487,and 0.9498,respectively,which are better than the existing methods.3.The MA detection method of fundus fluorescein angiography image based on multi-feature fusion is proposed.The saliency feature map of MA is obtained based on multiscale patch-based contrast measure,and the sparse representation feature by PatchImage model is extracted.Then the edge feature by SUSAN operator is obtained.Finally,the feature maps are fused at the pixel level to get the final detection result based on the similarity matrix.The sensitivity,positive prediction value and F1-score of the algorithm are 0.8133,0.9496,and 0.8762,respectively,which shows a quite good performance.4.The joint detection method based on multi-modal DR images is proposed.We extract the SURF feature vectors from the main vessel information of two image sources of the same patient and use the Euclidean distance as the similarity measurement to finish the acurate image registration combined with RANSAC measurement.Using the transformation matrix,we stitch and fuse the detection results,and present the MA lesions of patients comprehensively.The results show that our proposed method provides a good solution for the purpose of auxiliary diagnosis.