| Medical information processing is an interdisciplinary study that combines the study of medicine and the study of computer science.Fundus image processing and analysis is an important application of Computer Science in the field of medicine.Fundus screening based on the color retina images of the examined fundus can help detect the problems accurately and diagnose and treat the fundus disease in time.Currently,it is mainly through medical experts to check the related targets in fundus images,which makes it difficult to carry out a wider range of fundus detection.As a result,many patients would suffer visual impairment and even loss of sight due to the unavailability of timely detection and treatment.Therefore,utilizing the artificial intelligence methods such as machine learning,in-depth learning,pattern recognition and others to automatically detect related targets in color retinal images of the examined fundus,is of great significance for achieving large-scale fundus detections.With the access to abundant data and the guidance from some medical experts,based on clinical practice,the author conducts an exploration to attempt to solve the technical difficulties in fundus image processing such as target identification and classification under complex background;then a study of segmenting retina blood vessel,classifying retina arteriovenous blood vessel and identifying optic disc is also carried out.The main research findings are as follows:1.The automatic and precise segmentation of retinal blood vessels through fully connected conditional random field model based on multi feature fusion can help effectively detect the trunk and the end of blood vessels in the color fundus retinal image.Specifically,taking the advantage of the unique pattern of the strip-shaped and ribbon-like visible blood vessels with favorable local linear structure and using a combinatorial shift filter response model suitable for strip structure to extract the features of vascular trunk and vascular end.In view of the different characteristics of vascular trunk and vascular end,symmetric and asymmetric filtering models should be respectively used for tracking.Then both the symmetric and asymmetric responses of the combinatorial shift filter model and G channel pixel gray value can be used to co-construct an eigenvector library,hence to further improve the low contrast and segmentation of micro-vessels.2.A classification method of retinal arteriovenous blood vessels based on context related features is proposed.Due to the unique features of the arteriovenous blood vessels in the color retina images of fundus,the multi-scale context is represented as a probabilistic marking template with different ranges.At the local level,the pixel and the marker mapping describe the local image context of each pixel;at the regional level,a medium scale probabilistic marking template puts medium-range context restrictions on the markers in each region;Probabilistic marking template at the global level provides global constraints to the markers of the entire image.In this sense,the representation of context information involves not only the consistency of the objects in the image,but also the geometric relationships between the objects.Vascular morphology and topology features can be extracted on the basis of previous vessel segmentation attempts.Then the context correlation features that have more satisfactory discrimination can be introduced in different scales which contain shape,structure,relative location,context information,etc.Potential function of the CRF model can eventually be constructed by means of JointBoost classifier based on these features,to realize the training of labeled samples,hence to achieve precise classification of arteriovenous blood vessels in the retinal image.3.A video disk recognition method based on convolution neural network combined with conditional random field is then proposed.Firstly,use convolutional neural network(CNN)to pre-classify the color fundus image,and define the probability of members of the class as first-order potential function of conditional random fields(CRF)model.Secondly,use the linear combination of Gauss kernel function to define the two order potential function of CRF model and replace the common 4 or 8 neighborhood structure with full connection neighborhood structure.Thirdly,add regional constraints,use Mean-shift segmentation method to attain super pixel,modify the classification results by calculating the super pixel posteriori probability mean value,and encourage the consistency of the results of connected domains.The integration of CNN model and CRF model can help get the essential features of pixels.Spatial context information can make the classification more accurate.Added constraints can preserve the local information of the target.Finally,use the mean field approximation algorithm to extract the entire model,hence to achieve the accurate recognition of the optic disc in the color fundus retina image. |
