Small Sample Medical Image Recognition And Classification Based On Machine Learning

Medical assisted diagnosis based on machine learning is one of the current research hotspots.The rapid diagnosis of patients' diseases in clinical practice has also become extremely important,which directly affects the diagnosis,treatment and prognosis of the disease.However,in reality,problems such as small types,small quantities,and balanced samples of medical image data hinder the development of assisted diagnosis based on machine learning.This paper studies the identification and classification of small-sample medical images,and applies the algorithm used in the rapid diagnosis of OCT(optical coherence tomography,OCT)images of the fundus uveitis.This article mainly studies the following two aspects:First,use traditional machine learning algorithms combined with multi-feature fusion to achieve rapid screening for uveitis.This paper selects OCT images of healthy people and patients with uveitis respectively,extracts the morphological features of the images,the statistical features of gray difference,the gray gradient co-occurrence matrix and wavelet transform and other features.The features are serially fused,and then Lasso is used for feature selection.A variety of machine learning algorithms are used for classification research.The results show that the support vector machine(SVM)based on Medium Gaussian kernel function achieves 90.3% classification accuracy,and its area under the receiver operating characteristic curve(ROC)is 0.97,which is the model with the highest accuracy among the studied algorithms.Second,use the deep network-based transfer learning method to classify and recognize the OCT image of the fundus of uveitis.First,realize the transfer learning based on Image Net image data set.After the data enhancement,three pre-trained models(Res Net50,Inception v3 and Dense Net121)based on the Image Net data set were used to migrate to the target data set(uveitis fundus OCT image data set)for training and testing.Secondly,on the basis of the above content,the transfer learning based on the public fundus OCT image data set is realized.The three pre-trained models,Res Net50,Inception v3 and Dense Net121,were trained on the public fundus OCT image data set.In the process of model fine-tuning,we used three model fine-tuning methods.They are only training the fully connected layer,training the last convolution block and fully connected layer,training all networks.Finally,the models obtained by using different model finetuning methods are migrated to the target data set for testing.The experimental results show that the Inception v3 model achieved an optimal test accuracy rate of 92.4% when training the entire network,and the accuracy rate was greatly improved.This paper applies machine learning classification algorithm and migration learning algorithm to the actual classification of OCT images of the fundus of uveitis for the first time,and realizes the recognition and classification of small samples of medical images.It is an exploratory research on the diagnosis of uveitis.This is of great significance for the auxiliary diagnosis of uveitis.