Research On Liver Tumor CT Image Classification Based On SENet

In recent years,the number of tumor patients in China has been on a sharp rise.Among them,malignant tumor is a kind of disease that seriously endangers human health.If the disease can be detected in time and diagnosed early,it can effectively improve the cure rate of patients and lay the foundation for a good follow-up prognosis.Computed tomography is a common non-invasive medical diagnostic method for detecting malignant tumors,which can not only make a quick diagnosis for doctors but also distinguish undetectable lesions relatively easily.The continuous development of artificial intelligence has made it a trend to combine deep learning with computed tomography to assist physicians in diagnosis.Therefore,this thesis uses deep learning methods to classify CT images of liver tumors.1.To address the problem that the lesion area is more similar to its surrounding tissues,the contrast is low,and the information of lesion edge features is lost due to the influence of background noise;this thesis proposes a SENet liver tumor classification method based on multiscale feature extraction.This method first adds a hierarchical residual module,which is used to extract multiscale features of the image while expanding the network receptive field.Secondly,a dual attention feature extraction module is constructed to reduce the influence of redundant information and strengthen the model’s ability to identify the lesion areas.Then,a multi-branch feature extraction module is constructed in parallel with the atrous convolution to enhance the semantic information without losing the image feature information.Finally,octave convolution is used to replace ordinary convolution to reduce the number of parameters and improve classification accuracy.The method proposed in this thesis achieves the best performance under a variety of evaluation metrics,and the classification accuracy improves by 9.92% to 87.74% compared to the benchmark model.2.To address the problems of existing methods for medical image classification,such as complex image texture without good discrimination,few labeled high-quality datasets,this thesis proposes a SENet liver tumor classification method based on image reconstruction and feature fusion.Firstly,a variational autoencoder is constructed to generate rough images and then uses its output as the input of the dense perception generative adversarial network.Secondly,the dense block and residual structure are nested in the dense perception generative adversarial network,and the convolution is grouped then merged to improve the information flow in the deep network.Finally,the encoder and generator are used as feature extractors to perform feature fusion with the classification network,thereby improving the generalization ability of the model and the classification results.The results show that the method can expand the dataset better,and the accuracy is improved by 11.35% to 89.17% compared to the baseline classification model.3.The liver CT image lesion classification system was designed and developed to realize user management,data preprocessing,image reconstruction,and image classification functions.The network model proposed in this thesis was applied to the system,which can assist physicians in diagnosis,improve the accuracy of treatment,and lay the foundation for the construction of intelligent medical treatment.