Feature Fusion Based On Deep Learning And Its Application In Medical Images

Medical image analysis is the main way for medical staff to diagnose diseases.Based on the characteristics of OCT retinal data and Chest X-Ray pneumonia data,this thesis proposes a deep hybrid model based on convolutional network and transfer learning.When using the method of deep learning,the data volume of OCT retina and Chest X-Ray pneumonia data is not very large,this thesis proposes a new model based on transfer learning that can extract features from multiple basic convolutional neural network models,and the basic convolutional neural network models used in this thesis are VGGG16,ResNet50,and Inception V3.In this thesis,the data of the input model is firstly transformed,next the convolutional neural network is used to extract the advanced features,then the advanced features are fused,noise is added,and the fused features are further extracted,reduce the redundancy features by means of dimensionality reduction and sparsity limitation,and finally use the softmax function to classify.The parameter initialization of this model uses the Xavier method,and uses batch normalization to avoid back and forth oscillation of the training process,which can speed up the training;In order to reduce overfitting,use the dropout method and use LI regularization in loss function calculations;In order to improve the robustness,noise is added to the features that extracted from the pre-trained base model.The classification accuracy of the new transfer learning model in our thesis is 97.84%on the OCT retinal data,and the classification accuracy in the Chest X-Ray pneumonia data classification problem is as high as 99.17%,the classification accuracy in the liver cancer data is 98.3%.For practical application,this thesis also encapsulates an application for detecting medical images based on the proposed algorithm,which can assist the doctor to complete the image determination work.