Intestinal Polyp Recognition Algorithm Based On Feature Coding And Saliency

Polyp is one of the most common diseases in the human digestive tract and may lead to high-risk precancerous lesions.The timely detection and treatment for polyps plays an important role in preventing cancer and can save a huge amount of medical expenses.Endoscope is the main method to collect intestinal polyp images.However,due to the large number of images produced by endoscopy,the diverse shapes and sizes of polyps,the recognition effect of polyps mainly depends on the experience and skills of doctors,thus the artificial detection of polyps usually increases the burden of doctors and it may lead to missed diagnosis.Therefore,the study of automatic polyp detection is of great significance to relieve the pressure of clinicians and improve the reliability of diagnosis.The research contents and contributions of this thesis mainly include the following three aspects:(1)System design of intestinal polyp recognition based on feature coding and location based on saliency detectionThe existing auxiliary diagnosis system is usually a single image classification and lesion location.It is lack of consideration for specific scenes and diseases.In this thesis,an integrated system is proposed to integrate polyp image recognition and polyp region location in the intestinal environment.The system consists of two parts.The first part: the main information of the endoscope image is first extracted by using the feature and feature coding algorithm.Then the encoded features are pooled by max-pooling method,and finally the feature codes are input into the classifier to realize the dichotomy of polyp image and normal image.The second part: the target recognition algorithm based on saliency is used to generate the corresponding saliency map for the identified polyp images.The significance value of each block in the saliency map is further detected to assist the doctor in diagnosis.The system is capable of classifying polyps and locating polyp areas in a large number of intestinal endoscopy images.(2)Feature coding based polyp image classification methodAt present,the mainstream feature coding methods mainly focus on the objective function,however,it ignores the internal relationship between each base vector in the initial codebook.The design of codebook is worthy of further study because the final feature codes input to classifier is determined by codebook.In addition,our study have considered the effect of intestinal tubular structure on polyp image classification.Therefore,this thesis proposes an automatic polyp image recognition method based on the Salient Codebook Locality-Constrained Linear Coding(SCLLC)with Annular Spatial Pyramid Matching(ASPM).Firstly,the detailed texture features are extracted from the samples including normal image and polyp image,and the initial codebook is obtained by K-means clustering method.Secondly,the SCLLC algorithm is used to transform the initial codebook into the salient codebook,and the codebook is used to encode the features.Thirdly,in order to improve the efficiency of intestinal image processing,this thesis proposes a maximum pooling strategy based on spatial pyramid matching,which pools the computed feature codes.Finally,support vector machine(SVM)classifier is used to perform the polyp image classification task.The experimental results show that the algorithm achieves 94.10% accuracy and 91.20% sensitivity and 97.01% specificity in intestinal polyp image classification,which is superior to the existing mainstream methods based on feature coding.(3)Recognition of polyp region based on saliencyThe existing saliency detection algorithm usually uses the surrounding edge as the background prior and the image center as the target center prior,which is hard to recognize the background area of intestinal image correctly,and can not process multiple target positions.Therefore,this thesis proposes a saliency polyp detection algorithm based on the priori of intestinal center and contour background.Firstly,the detection method of intestinal center based on dark area and contour based on edge is proposed,which is used as a background priori.Then,the average geodesic distance is used to measure the similarity of different regions of the image,and the background region is expanded.Finally,a Gaussian center priori model of multi connected regions is proposed,which optimizes the significant image to get the final detection result.The experimental results show that in the final significant image comparison,the algorithm in this thesis can better detect the background area of the intestinal tract.Besides,the proposed algorithm can highlight multiple significant polyp targets effectively compared with the mainstream saliency detection algorithms.