Research On Mediastinal Lymph Node Recognition Algorithm Based On Deep Learning

The diagnosis of mediastinal lymph node metastasis mainly uses electronic computed tomography to detect the disease.The mediastinal lymph node metastasis is of great significance for the staging and dissection of lung cancer and other diseases.Medical imaging can currently obtain tomographic images,but the mediastinum structure is complex,and diagnosis requires clinicians to have rich experience and sufficient time.This paper solves the problems of complicated mediastinal structure,large differences in lymph node size,and wide distribution areas.We propose a two-step cascade mediastinal lymph node segmentation algorithm based on convolutional neural network and attention model.The algorithm improves the performance of mediastinal lymph node segmentation and detection by introducing medical priors and enhancing channel,spatial,and multi-level features.The main work is as follows:This article first collects,organizes,and expands the mediastinal lymph node dataset,and marks the mediastinal organs and lymph nodes.Due to the complex structure of the mediastinum region,a two-step segmentation step was formulated based on medical prior knowledge.The first step is to divide the mediastinal organs and lymph nodes.In the first step,the segmentation result and the mediastinal organ is used as a mask,and the area of interest of the lymph node is obtained through opening and closing operations and filling holes,which can greatly reduce background interference.In the second step,a variety of attention mechanisms and multi-level feature modules are used to solve the problems of widespread lymph node distribution and large size differences.In the second step,for sequence images,the original two-dimensional global aggregation module and the dual attention module are modified into three-dimensional modules,and the adaptive receptive field is introduced into the dual attention module according to the large difference in lymph node scale.Because the global aggregation module occupies a large amount of memory and the recognition effect of small lymph nodes in the model is not good,a global context module and a feature fusion module are designed,and a mediastinal lymph node segmentation network is proposed.The global context module can calculate the weights between voxels of feature blocks and enhance the features of lymph nodes at different locations;the feature fusion module combines the edge texture features of shallow convolutions and the abstract semantic features of high-level convolutions to reduce small targets caused by convolution operations Lost.Experimental results show that the Dice Score of this method can reach 76.08%,which is better than other mediastinal lymph node segmentation algorithms.This article analyzes the main problems of mediastinal lymph node detection,proposes a two-step cascade method based on medical a priori,and modifies and proposes different attention mechanisms and feature fusion modules based on the mediastinal sequence images,which have achieved good results.Clinically,it can assist physicians to improve the efficiency of diagnosis and treatment.