Lung Nodule Recognition And Classification Based On CT Images Of The Lungs

At present,in the clinical diagnosis and treatment of lung diseases,doctors still need to manually outline the lesion area.The manual screening will be affected by the professional ability and subjective experience of doctors to a certain extent,and the test results of each tester have hundreds of images,which makes the workload of testing doctors very large.In this study,with the help of computer-aided doctor detection of pulmonary nodules,the experiment of pulmonary nodule recognition,segmentation,and classification was carried out based on a deep neural network.The specific research work includes the following aspects:First,preprocess the data.LUNA16 is used as the experimental data.The label marking specification of this data set is authoritative,and the amount of data can meet the requirements of the deep learning model.In the process of pulmonary nodule recognition and segmentation,firstly,the image mask file is generated according to the original data,and then the CT image is analyzed to obtain the slice thickness and window width.After the window level,the image denoising is carried out,and then the data is interpolated and sampled.The sampled data is stacked to generate a 3D pulmonary nodule slice image.The preprocessing of pulmonary nodule recognition and segmentation data is basically completed.In the classification of pulmonary nodules,due to the large difference in the number of positive and negative samples in the data set,the positive samples are enhanced by 40 times to obtain the positive samples,and then 20% of the non-pulmonary nodules are randomly sampled to obtain the negative samples.The pre-processed positive and negative samples are divided into 80% training data and 20% test data.Secondly,this study proposes an improved 3DVNet model for lung nodule recognition and segmentation.The structure of the improved 3DVNet model is simpler.The improved network includes 4 layers of convolution layer and 4 layers of pooling layer,and the corresponding four anti convolution layers and 4 upper sampling layers are reserved.The intermediate convolution layer is added to the extracted and restored features.The convolution layer and deconvolution layer of the feature map of the same size are connected by the cascade layer,which can save a lot of computing time and release more computing resources.At the same time,it retains the characteristics of the VNet model with a large receptive field and alleviates the overfitting problem of the original network.The experimental results show that the segmentation accuracy is 88.29%(DSC value)and the predicted image intersection to union ratio is 88.25%(IoU),which is better than most of the existing methods in the literature.The IoU of segmenting various morphological nodules by this method is 86.87% at the lowest and 92.85% at the highest.The recognition and segmentation effect is good,which proves the universality and robustness of the improved model.Finally,this paper proposes an improved VGG model for classification.The improved VGGNet consists of ten convolution layers and two full connection layers.The original VGGNet algorithm contains three full connection layers,and the improved VGGNet reduces one full connection layer.The results show that removing one full connection layer has no obvious impact on the network effect,and can effectively reduce the number of parameters.The loss function obtained from the training of the improved VGG model shows good convergence.The prediction results can be obtained by using the statistical results of the confusion matrix.The prediction results are 17357 true negatives(TN),16 false positives(FP),431 false negatives(FN),and 12196 true positives(TP).The results can evaluate the network performance.The overall accuracy of the network is 98.51%,the false positive rate is 0.921%,and the missed detection rate is 3.413%.From the calculation results,it can be concluded that the improved algorithm has a better classification effect.