Key Technologies Of Medical Images Auxiliary Diagnosis Based On Depth Vision

The lungs are a pair of sponge-like organs in the human body that play the most important role in respiration.Given the importance of the lungs,lung diseases have also received a great deal of attention from the clinical medical and computer science communities.The diagnosis of medical images requires specialized and experienced physicians,but there are few resources for such physicians in hospitals,and simply screening and discriminating the large number of medical images by hand can lead to errors in judgment due to overwhelming fatigue.As the development of deep vision technology becomes more and more mature,the application of this technology to medical images to develop medical image intelligence-aided diagnosis system has become a hot research topic,and the State Council has called for its development in the Development Plan of New Generation Artificial Intelligence.Although hospitals keep a large amount of medical image data needed for training and learning of deep vision techniques,the special characteristics of medical image datasets,which are different from natural images,lead to privatization,miniaturization,uneven structure and lack of multiple classifications.In order to solve the shortage of large medical image data sets and improve the accuracy of medical image recognition by computer-aided diagnosis system,this thesis takes six types of chest CT images as an example.The main research contents are as follows:(1)In this thesis,a large structured and balanced chest image dataset Chest-6CT is created,which has a total of 100987 internal sample data,including 6 chest CT images of 16789COVID-19,16715 Common Pneumonia,16798 Lung Squamous Cell Carcinoma,16794 Lung Nodules,16966 Non-small Cell Lung Cancer and 16886 Normal lung cancer,are compared with the existing dataset,which eliminates the defects of privatization,miniaturization and unbalanced structure of the current medical image dataset used for medical image assisted diagnosis technology development work.(2)In this thesis,we propose to complete the extraction of lung parenchyma within CT using a segmented SEG model,which consists of an adaptive thresholding algorithm and a flood filling algorithm.The redundant information on CT other than lung organs is removed using the SEG model,eliminating the difficulty of the model to extract lung lesion features from a non-pixel-level labeled full-size lung CT image dataset.The SEG model is experimentally validated for segmentation on the dataset Chest-6CT,and its Dice similarity coefficient reaches 0.972,which can be used for lung parenchyma segmentation,and the model is comparable to different The segmentation performance of the SEG model has the best performance when compared with different segmentation models on the same dataset.(3)SEG-VIT model is proposed to solve the problem of low accuracy of existing auxiliary diagnosis models,and the classification performance of the model is verified by experiments on the data set Chest-6CT.The model is developed based on the Vi T model and supplemented by the SEG segmentation model,in which Vi T performs better than the existing convolution neural network in natural image classification.The training samples first extract the lung parenchyma through the SEG segmentation model,and then enter the classification network training to get the chest CT image classification model SEG-VIT.The experimental results show that the Accuracy,Precision,Recall,Specificity and Full1 of chest CT images are 0.975,0.973,0.971,0.994 and 0.972 respectively,and the training time is greatly shortened.At the same time,SEG-VIT model and different models are compared on the same data set,SEG-VIT still shows excellent classification performance.Finally,the focus area diagnosed by the model is located and visualized with the help of visualization tool CAM to further meet the actual needs of doctors’ diagnosis.