| Medical imaging and its related analysis technology have been widely applied in medical field,and gradually become significant basis for medical diagnosis and treatment.As key of auxiliary diagnosis and treatment,it has important research and clinical value.Retinal image is generated by fundus camera and is applied as diagnosis image in ophthalmology,and retinal image analysis will find the variation of vessel structures,and provide auxiliary information for diagnosis and treatment of retinopathy caused by diabetes and hypertension etc.Because of the complex feature of retinal images,few research achievements have been obtained in automated process and analysis technology for real applications,there are some key technical bottlenecks need to breakthrough.In this paper,focus on retinal image preprocessing,optic disc positioning and segmentation and vessel segmentation etc.,the following research work is carried out.1.Optic disc in retinal image has fuzzy edges,and the vessels are thin,because of noise,nonuniform illumination,poor contrast between target and background etc.,it is hard to precisely segment them.This paper proposed specialized image denoising method for retinal image.The traditional denoising technologies process the whole image globally,while smoothing the image noise,the target edge details are also smoothed out,which will cause the inaccuracy and incomplete of the segmentation results.So this paper proposed local adaptive noise filtering(LANF)method,centered at each pixel,a neighbor area is generated,and the intensity status of the area is analyzed,if the intensity difference of the neighbor area is within a certain range,then the pixels are considered to be the same class,if the intensity difference between the center pixel and the average intensity of the neighbor is higher than the threshold,then it is considered as noise,and its color value is replaced by the average color of the neighborhood,otherwise no denoising is performed.According to the neighbor context,LANF can determine whether to smooth a pixel or not,which can effectively remove the salt and pepper noise,while remaining the edge details.2.This paper proposed automatic optic disc(OD)localization and accurate segmentation method based on intensity and mathematical morphology.Because of fuzzy edges,nonuniform illumination etc.,it is hard to localize and segment OD accurately.This paper firstly analyzed the intensity distribution status of the image,and for each image,an intensity threshold was adaptively determined,and those pixels whose intensities were higher than the threshold were marked as OD candidates,according to its normal size,a structural element was generated and used to localize OD based on erosion mechanism.Centered at it,a minimum enclosing circle which can completely include OD was determined,and the internal vessels were removed.By learning the intensity features of OD edges,those pixels conform to the features are marked as contour which can segment OD accurately.Compared to the traditional active contour model,this method doesn't need manual intervention,which is more efficient and accurate.3.This paper proposed retinal image vessel segmentation method based on pixel specificity.Based on pixel intensity,pixel specificity is defined and computed for each pixel,and according to the definition of pixel specificity,if a pixel has larger pixel specificity value,then it is more probable to be vessel,otherwise it is more probable to be background.Take this theory as foundation,for different types of retinal images,two vessel segmentation methods were respectively proposed.(1)Retinal vessel segmentation method based on local adaptive pixel specificity threshold.Due to nonuniform illumination on retinal image,the intensity status of different regions may have large difference,if the vessels are segmented according to the same standard,then the segmentation accuracy may be reduced;and if there are lots of salt and pepper noise,the segmentation of small thin vessels will be affected.Aimed at these problems,firstly LANF was applied to remove noise,then a pixel specificity threshold was set to pre-segment the vessel,and the image was divided into nonoverlapping 16?16 patches.For each patch,local pixel specificity threshold was computed adaptively by gradient decent,which was then used for vessel segmentation within each patch.Finally the vessels segmented from each patch were combined together and got the whole segmentation result.By denoising,collective awareness to search the missed vessels and broken vessel segments linking,the whole vessel structure was achieved.(2)Retinal vessel segmentation based on self-adaptive classification strategy.Different retinal images may have different intensity status and other structural features,e.g.for some images,the optic disc may have larger intensity values,and their edges are easy to be segmented as vessels;some images may have poor contrast,the edges of some vessels are fuzzy,and some vessels have central light response;some images have many lesion regions which have low intensity values and are easy to be segmented as vessels etc.For the problems mentioned above,the segmentation process was divided into 3 phases to implement.Firstly according to the relationship of pixel specificity and vessel probability,a higher pixel specificity threshold was determined,and those pixels with higher pixel specificities were extracted as main vessels,and then self-adaptive pixel classification strategy by multi-agent was applied,where each undetermined pixel acted as an Agent,within the range of multi-scale pixel specificity thresholds,the agent revised its status according to the status of its neighbor agents,and gradually completed the pixel classification.Finally the noise was removed by double-window denoising and the whole vessel structure was achieved.4.This paper designed and trained a convolutional neural network(CNN)which can implement the parallel segmentation of optic disc and vessel in retinal image.Before segmentation,the retinal images were normalized to make sure that the background intensity and contrast of all the images are consistent.For each effective point in retinal image,three neighbor regions with different sizes were extracted as three channels input for the CNN and were feedforward through 5 layers.The output layer has three neurons which correspond to three classes respectively: background,optic disc and vessel.The CNN was trained and evaluated on the publicly available DRIVE database,and the average segmentation accuracy was 92.99%,while the highest accuracy for a single image is 94.95%,and the lowest accuracy is 88.79%.This paper used a CNN to implement the accurate segmentation for optic disc and vessel of retinal images simultaneously,compared with other CNNs,it reduced the training time and is more efficient. |
PDF Full Text Request