Research On Medical CT Image Recognition Based On Deep Learning

Lung cancer is the leading cause of cancer death in humans,which has posed a huge threat to the life and health of people.In clinical trials,early intervention of lung cancer plays a vital role in improving the survival rate of patients.However,most of the lung nodules have complex and diverse forms,some malignant nodules have little difference from benign nodules in the early stage,which results in easy missed diagnosis and misdiagnosis in lesion screening,and increases the detection difficulty and work burden of radiologists.At present,the lung nodule diagnosis algorithm based on 3D convolutional neural network still faces the challenges of low accuracy of nodule segmentation and benign and malignant classification,as well as high model complexity.To solve the above problems,this paper mainly carried out the following research:(1)In order to achieve accurate localization and fine edge segmentation of nodules,this paper proposes a 3D Rem-UNet(3D Residual connection Multiple branches Hybrid Attention UNet)lung nodule segmentation model based on 3D UNet segmentation network.In this model,a residual connection module is introduced into the 3D UNet coding path to enhance the utilization rate of the underlying features and avoid the gradient disappearing in the training of the network.At the same time,Multiple branches Hybrid Attention(MHA)module is designed and proposed to achieve adaptive weight allocation of channel and spatial features,and further enhance the ability to extract key image features.Then the group normalization algorithm is used to avoid the influence of small batch data on the statistical estimation.Finally,the proposed pulmonary nodule segmentation method is verified on the LUNA16 dataset.The experimental results show that the Dice similarity coefficient of 76.42% is obtained by this method,which is4.73% higher than that of the baseline network,and the segmentation accuracy of pulmonary nodules is effectively improved.(2)For ensure the accuracy of classification of pulmonary nodules benign and malignancy and achieve more lightweight the demand of the network structure,a classification algorithm of lung nodules(3D Depthwise Separable Attention ResNet,3D Dsa-ResNet)is proposed based on 3D ResNet18 network model.The 3D Dsa-ResNet model uses the Stage module as the feature extractor,which widens the network width by adding depth separable convolution in parallel to the residual learning module to avoid the problem of single feature extraction and feature redundancy cause by simple stacked network modules.At the same time,an improved attention mechanism is introduced into the model to further improve the classification accuracy of pulmonary nodules.In addition,in view of the imbalance of good and bad samples in the data set,image enhancement technology is used to expand the samples,which can effectively enhance the diversity of data.In order to verify the validity of the model,LUNA16 data set is selected for experiments,and the number of parameters and floating point calculation amount of the model are analyzed.Experimental results show that the proposed 3D Dsa-ResNet pulmonary nodule classification method could achieve 94.57% accuracy,and compared with3 D ResNet18,the number of parameters and calculation amount of the model decreas by83.54% and 35.49%.Compared with the existing image segmentation and classification models,the experimental results show that the proposed segmentation and benign and malignant classification methods can effectively improve the segmentation effect and classification accuracy of pulmonary nodules.Furthermore,the validity and feasibility of the proposed segmentation and classification algorithm are verified,which has certain research significance.