Research On Detection Technology Of Lymph Node Metastasis Based On CT Images

Accurate detection of lymph node metastasis is a prerequisite for cancer staging,and the gold standard for detection is biopsy of lymph node tissue.However,this method causes additional trauma and complications to the patient and increases the risk of cancer metastasis.In addition,the high cost and difficulty of sampling limit the applicability of this method.With the rise of deep learning,more and more researchers are exploring the use of convolutional neural networks to mine lymph node features in computed tomography(CT)images for classification.However,there are difficulties in predicting lymph node metastasis base on CT images.First,traditional convolutional networks use fixed-size input images,and the large variation in the volume of different individual lymph nodes can be problematic when resampling the images.Second,the contrast between lymph nodes and surrounding tissues is low,and features are difficult to be mined.Last,the common problem of insufficient data in medical field can limit the performance of deep learning methods.Therefore,this thesis conducts research on lymph node metastasis detection method based on CT images.Firstly,a multi-scale feature fusion network is proposed for the problem of large variation in lymph node size.Then a multiple image transformation semi-supervised network is proposed to solve the problems of low contrast of lymph nodes and insufficient amount of data.Finally,based on the above research,this thesis designs and implements a medical assistance system to provide assistance to doctors in clinical diagnosis and effectively improve their work efficiency.The work in this thesis includes the following three aspects:(1)In this thesis,a multi-scale feature fusion network is proposed for the detection of lymph node metastasis.Multi-scale inputs can solve distortion problems caused by excessive resampling multiples.A feature fusion module based on self-attention mechanism is used to fuse the features of each branch and fuse the local features.Considering that the spatial distribution of features at different scales is similar,a feature consistency loss function is proposed to constrain the similarity of features of each branch.In addition,the entire network is a fully convolutional structure,which reduces the amount of parameters of the network,and classification from multiple dimensions also enhances the robustness.(2)In this thesis,a multiple image transformation semi-supervised network is proposed for the detection of lymph node metastasis.Image transformation with random intensity is used to map shape features into multiple feature spaces,thereby enhancing the network’s mining of shape features.At the same time,the randomness of image transformation intensity masks the differences between different datasets,so that the unlabeled dataset can be better used for semi-supervised training,allowing the network to learn more lymph node features and improve the accuracy of prediction.(3)Based on the above research,this thesis designs and implements a medical assistant system for lymph node metastasis detection based on CT images.The system supports the functions of inputting CT images in various formats,getting patient information from database,visualizing CT images,metastasis prediction,and saving reports.It can assist doctors in diagnosis and greatly improve the efficiency.Taking mediastinal lymph nodes in CT images of non-small cell lung cancer as an example,this thesis conducts detailed experiments on a dataset of 148 patients with lymph node metastasis labels,and the model achieved 92.37% accuracy.The TCIA public dataset containing 90 patients was introduced for semi-supervised training,and achieved 92.45%accuracy.Both methods are better than the comparison methods,which proves the effectiveness of the algorithm in this thesis.