Study On Deep Learning Based Benign And Malignant Classification Technology Of Pulmonary Nodules

Lung cancer has the highest morbidity and mortality in the world,posing a huge threat to human life and health.In the diagnosis and treatment of lung cancer,computer-aided diagnosis technology is urgently needed to assist doctors to classify benign and malignant pulmonary nodules.Traditional methods automatically diagnose pulmonary nodules through three steps: segmentation of pulmonary nodules,feature extraction and classification.However,these methods are complicated in designing and low in classification accuracy.Among the methods based on deep learning,most methods design two-dimensional convolutional neural networks(CNN)to automatically classify pulmonary nodules and have achieved good results,but these methods inevitably lose the three-dimensional spatial information of CT images;The methods based on 3D CNN consumes a lot of training resources,which makes it difficult to train the network.Considering the existing problems,this thesis innovatively proposes a multi-dimensional information fusion network designing method,which optimizes the classification performance of benign and malignant pulmonary nodules by fusing the information extracted from two-dimensional and three-dimensional pulmonary nodule images,and reduces the training parameter requirements.Based on the method,this thesis proposes a Multi-Dimensional Fused network(MDF-Net)based on decision-level fusion and a Multi-output Multi-Dimensional Fused network(Mo MDF-Net).In MDF-Net,pretrained Dense Net with shared parameters are used to fine-tune layer by layer to extract features from two-dimensional pulmonary nodule slices on different perspectives,and a three-dimensional Dense Net is designed from scratch in three-dimensional sub-network to extract features from three-dimensional pulmonary nodule slices on transverse section.The network uses an attention module to fuse information from different dimensions at decision level and output the final prediction results.In Mo MDF-Net,multi-scale features of different sub-networks are extracted at input level to improve the classification performance.Besides,at different stages of the network,feature maps,which containing different levels of semantic information,are fused and fed into different classifiers to output intermediate predictions,alleviating the problem that deep networks are prone to fall into local minima.The final predictions are obtained through averaging the intermediate ones.Experimental results show that the proposed network designing method of fusing information can effectively solve the problem of classifying benign and malignant pulmonary nodules,and achieve better results than the single 2D and 3D neural network design method.Besides,compared with 3D convolutional neural networks,the proposed network needs fewer training parameters.