Research On Medical Image Classification Methods Based On Deep Learning And Attention Mechanism

Image classification technology is a crucial research direction in the field of deep learning.This thesis applies a deep neural network to diagnose and identify medical images.And the principal purpose of this thesis is to improve the classification performance of the network based on the traditional convolutional neural network(CNN).In this thesis,specific studies on the Rhizoma Paridis microscopic and ulcerative colitis datasets are carried out.In identifying microscopic images of Rhizoma Paridis,this thesis adds different attention mechanism modules to the image based on the CNN model.The principle is to imitate the human eye to observe the object and lock the meaningful area after scanning the overall image.Its core mechanism is to focus on the critical information in the image by changing the weight of different feature learning according to the loss change during model training.Major improvements include:1.When enhancing the microscopic image of the Rhizoma Paridis,the Mixup method is added to improve the model’s generalization ability by adding disturbance.2.To improve the ability to extract features during network training,add an attention mechanism to the model to improve the discrimination effect of the microscopic images of the Rhizoma Paridis.On the Rhizoma Paridis microscopy dataset,the experimental results show that compared with the ResNe Xt101 model,the four accuracies,precision,recall,and F1 score indicators have increased by 1.73% and 2.10%,1.83% and 1.91%,respectively.The experimental results show that using the convolutional attention module combined with the backbone network has the highest performance,which is beneficial to the extraction of image features.In the diagnosis of ulcerative colitis images,due to the limited feature extraction capability of traditional convolutional neural networks,a combination of advanced image features learned by the hierarchical architecture of the CNN model and spatial dependencies of image regions learned by the Recurrent NeuralNetwork model is proposed to enhance the ability to improve the performance of image classification.Major improvements include:1.A neural network called UC-DenseNet,which combines CNN and RNN,is proposed to replace traditional ulcerative colitis image recognition methods.2.An improved attention mechanism is proposed,which can adaptively determine the kernel size of the network model through a function of the channel dimension.At the same time,to further determine the position of the key features of the image,the spatial attention module is used to lock the detailed information in the image and extract it and suppress the remaining invalid features.3.Summarize the performance improvement of the network after augmenting the test data and explore the impact of using different kernel sizes in the attention module on the network performance.The experimental results show that when UC-DenseNet diagnoses endoscopic remission,the five indicators of accuracy,precision,recall,F1 score,and area under the curve are improved by approximately 1.08%,1.06%,1.81%,1.06%,and 0.94%,respectively.In addition,when UC-DenseNet diagnoses the degree of endoscopic disease,the four indicators of accuracy,precision,recall,and F1 score are improved by approximately 2.53%,1.74%,4.29%,and 3.54%,which is better than the traditional CNN classification method.The results show that the proposed method shows promising performance on three different colonoscopy datasets.