Research On Medical Image Classification Algorithm Based On N-Gram Model

Medical image classification technology is a key technology in computer aided diagnosis.However,the 'semantic gap',data unevenness,and complex dimensions have hindered the application of classification models in clinical.As a classic language model,the N-Gram model shows superior performance for solving such problems.This model is simple and efficient,which can extract the abstract semantic features of text based on the principle of statistics while taking into account the context information of local features.In this paper,the application of N-Gram model is extended from traditional text processing to medical image classification by using learning transfer method,and the transition algorithm is studied deeply.Firstly,the source image is encoded by gray-scale function.Secondly,N-Gram visual corpus of samples is collected by sliding window.On this basis,significant N-Gram visual corpus is selected as the feature of samples by using constraints and significance test,whose frequency vector is the feature vector of samples.Finally,a classifier is used to classify the image targets.In order to test the performance of the algorithm,two tasks are completed in this paper: the benign and malignant classification of thyroid nodule ultrasound image and the identification of dry macular degeneration of optical coherence tomography(OCT)image.In order to reduce the sensitivity of the algorithm to noise,the preprocessing process is simplified and improved according to the characteristics of focus imaging.The improved algorithm based on the CV-RSF model is used to complete the nodule segmentation in task one.The improvement based on the anisotropic diffusion model is used to complete the speckle noise removal processing in Task Two.In order to test the algorithm's dependence on the classifier,Fisher discriminant,Support Vector Machine(SVM)and BP neural network are used for comparative experiments.Finally,the highest accuracy of them in Task One are 97.06%,100% and 100% respectively,and the highest accuracy in Task Two is 99.61%.The experimental results show that the algorithm proposed in this paper has excellent classification performance,strong generalization ability,high robustness and low dependency on the classifier,which provides a set of efficient technical methods and research ideas for researchers to build computer diagnostic models.