| With the increase of people’s attention to their own health and the progress of medical equipment,there are countless images in the medical field,but because most of the images lack the necessary labels,the medical resources can not be used effectively.If these resources can be used effectively,it will be of great help to the research in this field.At present,the diagnosis of lung cancer mainly depends on the doctor to find the pulmonary nodules in the image,and then diagnose them.Due to the rapid increase in the number of images and the size and shape of pulmonary nodules,which are very similar to the blood vessels of the lung,it is very difficult for doctors to diagnose.In order to solve these problems,this paper uses the CT image of lidc-idri as the data set,studies the lung nodule segmentation algorithm based on improved U-net,aiming at the low efficiency and difficulty of manual lung nodule segmentation,and studies the multi label lung nodule segmentation algorithm based on semi supervised learning,aiming at the problem that it is difficult to train experienced doctors and the lack of complete labeled data sets Therefore,this paper is divided into two steps to introduce: the first step is to segment the pulmonary nodules in the lung CT image;the second step is to generate multiple tags from the segmentation results obtained in the first step.Through these studies,to complete the diagnosis of pulmonary nodules.The two main research points are the segmentation of pulmonary nodules in lung CT images and the extraction of calcification,lobulation and other features of segmented pulmonary nodules for multi label generation.For the first step of lung nodule segmentation,due to the low segmentation accuracy and insufficient data volume when segmentation or feature selection is done manually,the u-net network which is often used for this kind of dataset segmentation is used for segmentation.Because of the gradient loss and low feature utilization,the network structure is improved to improve the efficiency of lung nodule segmentation Segmentation accuracy.The residual connection module is added to improve the gradient problem of the model;the common convolution in the model is replaced by the expansion convolution block to improve the utilization rate of features;the attention gating module is added to filter the features in the model down sampling to reduce the redundant features in the model,and then the output of each layer is fused.When the fusion is completed,the spatial and channel annotation is used to complete the segmentation of pulmonary nodules.The experimental results show that the DSC and PPV of the improved model are 83.52% and 82.31%,respectively,which can effectively improve the segmentation accuracy.For the second step of multi label generation of pulmonary nodules,the pulmonary nodules and their text description are usually put into the neural network for training.However,due to the small number of pulmonary nodules in the data set,the semi supervised method is used for network training,and the nodules obtained by the first step segmentation and the corresponding text description are put into the network for semi supervised training Therefore,a pyramid layer is added in front of the first fully connected layer in the model to fuse information of different scales.At the same time,an attention layer is added before the last fully connected layer to improve the attention of the target area.Through comparison with relevant experiments,the average accuracy of this model for eight different characteristics of pulmonary nodules was 85.32%,78.48%,86.13%,80.34%,62.46%,77.15%,86.54%,85.07%,respectively,which proved that the above improvements could obtain better experimental evaluation indexes.Finally,according to the above research,an auxiliary diagnosis system is developed.Through the above two steps of operation on the input CT image,the system finally generates the description text of the image and displays it to the user. |
PDF Full Text Request