| Most diagnosis of laryngoscope diseases is through the collection of laryngoscope endoscopic images,combined with the doctor’s direct observation.However,it is an extremely arduous work on doctors’ part to identify abnormalities in medical images with the naked eyes alone.Computer-aided technology has been put into use in some medical fields.However,there has been little research on laryngoscope images.This thesis studies the laryngoscope image data of otolaryngology head and neck surgery in a hospital in Jilin Province.The laryngoscopy images of normal and reflux patients are classified for three methods,and the subtle differences between the two groups are magnified to improve the diagnostic efficiency and accuracy.In the first part,the entropy weight method K-means-GMM(Gaussian Mixture Model)algorithm is used to classify the laryngoscope image data.Considering the influence of the sample index on K-means clustering,the Euclidean distance is weighted by the entropy weight method,and the clustering result obtained by the entropy weight method K-means is used as the initial parameter of the GMM classifier.Use this algorithm to classify laryngoscope image data,and the accuracy rate reaches 90.34%,which is 0.60% higher than the accuracy of the K-means-GMM algorithm.In the second part,the classification of laryngoscope image data is carried out by using the K-means-CNN(Convolutional Neural Network)algorithm of the improved activation function entropy weight method.First,a new activation function SRe LU(S-shaped Rectified Linear Activation Unit)is proposed,and then combined with the advantages of unsupervised learning and supervised learning.Use entropy weight method K-means in the unsupervised learning part and CNN(SRe LU)in the supervised learning part.The final accuracy of laryngoscope images reaches 96.07%,which is 0.33% higher than the accuracy of the Kmeans-CNN(Re LU)algorithm.In the third part,using the improved entropy weight method K-means-LeNet-5algorithm to classify the laryngoscope image data.Using a combination of unsupervised learning and supervised learning,in the unsupervised learning part,the entropy weight method K-means is used,and in the supervised learning part,the LeNet-5 network is improved,and the SRe LU activation function is used to enlarge it as much as possible The difference information of laryngoscope images,the accuracy of the algorithm for laryngoscope images is 96.35%,which is 0.31% higher than the unimproved LeNet-5(Re LU)network. |
