Medical Image Classification Technology Based On Deep Learning And Spatial-Frequency Feature Fusion

In recent years,a group of deep learning technologies led by convolutional neural network technology have emerged in the field of natural image processing.Especially in image classification tasks,deep learning methods have surpassed the effects of traditional methods.However,unlike natural images,the characteristics of objects in medical images may not be obvious,and there are similarities between the characteristics of different objects.Therefore,if a deep learning model designed for natural image classification task is directly used to classify medical images,its effectiveness will be reduced.In addition,most models based on convolutional neural network technology only focus on the spatial domain data of the image,while ignoring the frequency part of the image,which means the model do not utilize all the information provided by the image.In order to solve the above problems,this thesis proposes to use multi-scale features and fused spatialfrequency domain features to improve the classification accuracy of the model on medical images.The specific research contents include the following three parts:Research on the fusion algorithm of medical image semantic features based on attention mechanism.In this thesis,the multi-label classification method of fundus lesions is firstly studied,and an image feature fusion module is proposed.The high-level features with rich semantic information extracted by the backbone network are used to guide the refinement of low-level features of the network,so as to extract the multi-scale features of the image.According to the characteristics of the data,a lesion correlation learning module is proposed to learn the implicit correlation between image lesions to improve the classification accuracy of the model.By analyzing the experimental results,it is proved that the multi-scale features and the implicit correlation between the lesions in the image can improve the classification accuracy of the network.Research on spatial-frequency domain feature fusion medical image classification algorithm based on discrete cosine transform.By using the frequency domain data computed by discrete cosine transform,this thesis proposes a frequency domain information screening module based on attention mechanism for filtering frequency domain data.In addition,this thesis fuses the filtered frequency domain data with the spatial domain data provided by the backbone network to obtain the final spatialfrequency domain features.Through experiments on the glaucoma fundus image dataset,the conjecture that multi-scale features and spatial frequency domain features can improve the classification accuracy of the network is confirmed.In addition,this thesis presents a part of feature maps of the relevant frequency domain data.Research on spatial-frequency domain feature fusion medical image classification algorithm based on non-subsampling contourlet transform.In order to solve the problem that the method based on discrete cosine transform can only save part of the texture and position information of the image,this thesis proposes a multi-scale frequency domain feature extraction module based on the non-subsampling contourlet transform.This module extracts multi-scale frequency domain features from the transformed data with rich texture information and different scales.Then,the frequency domain features extracted by this module is fused with the spatial domain features extracted from the backbone network to obtain the final spatial-frequency domain features.Furthermore,this thesis conducts experiments on multiple public datasets and analyzes the experimental results.The experimental results prove that multi-scale features and spatial-frequency domain features can improve the classification accuracy of the backbone network.