Research On Object Detection And Intelligent Analysis Diagnosis Methods For Fundus Image

Medical image processing and analysis is to use a computer to automatically process,extract,and classify medical images.Due to the complex structures in the fundus images,there are still many key technical difficulties in the existing approaches,such as object localization under uneven illumination and low contrast,accurate object extraction under complex background,and the class imbalance problem in classification,etc.Based on the analysis of domestic and overseas researches,a series of effective approaches have been proposed in this dissertation for intelligent analysis and detection of the objects in the fundus images,including optic disc,blood vessels,macula,and lesions,achieving the intelligent diagnoses of both diabetic macular edema and diabetic retinopathy.The main works and innovations of this dissertation are as follows:Since the object detection in fundus images always suffers from uneven illumination,and complex background structures,an object detection approach based on visual local feature is proposed.This method uses visual saliency method to extract the object candidate regions,improving the computational efficiency of the algorithm.Then,local features are extracted from candidate regions and a sparse dictionary is constructed.On this basis,the use frequency of dictionary atoms is regarded as a "spectrum" feature for classification.This method has achieved high precision in the application of optic disc detection.According to the fovea detection always suffers from the problems caused by similar non-object regions and high computational load,this dissertation develops a novel fovea detection approach based on image prior.First,morphological approaches are employed to extract the blood vessels according to the characteristics of objects in the fundus images.Then,the optic disc center is determined by the prior knowledge of the blood vessels structure inside the optic disc.In order to reduce the influence caused by non-object regions,the region of interest of the macula is extracted using the prior information of the position between the optic disc and the macula.Meanwhile,structural features are extracted within the region of interest achieving automatic detection of the macula.The experimental results show that this method can locate the center of the macula accurately and efficiently.An unsupervised object detection method based on statistical analysis is proposed to solve the class imbalance problem in the traditional classification methods.This method employs sparse principal component analysis to construct the model,avoiding the problem of incomplete training caused by the difference in the number of samples under different categories.Then,an unsupervised classifier is designed to automatically identify the object.This method explores the intrinsic properties of the samples and detects the object through the statistical analysis method,which has achieved good results in the detection of microaneurysms.Considering that object detection in the fundus images always suffers from the problems of edge information loss and non-object regions interference,an object detection method based on superpixel multi-feature is proposed.The proposed approach first uses global and local contrast enhancement method to improve the contrast and brightness of the fundus image.Then a superpixel segmentation method is used to segment the image,obtaining the candidate regions with the edge structure information.Meanwhile,extracting their visual features,and a linear discriminant analysis method is used for classification.In order to reduce the misdetection rate caused by the non-object regions,this dissertation adopts the post-processing method to remove the influence caused by non-object regions achieving accurate detection.This method is applied to hard exudates detection,and compared with other similar methods,it can not only achieve high-precision but also maintain the edge information of object.Since the manual screening has a strong subjectivity,and the possibility of misdetection and missed detection,two intelligent diagnosis methods namely diabetic macular edema and diabetic retinopathy are proposed by combining the aforementioned intelligent detection approaches.The experimental results indicate that the proposed approaches can significantly improve the accuracy of fundus diseases diagnosis.