Computer Aided Diagnosis Research Of Lung Tumor PET/CT Based On Convolutional Neural Network

Background Deep learning is introduced into the field of machine learning to make it closer to the original goal of artificial intelligence,convolutional neural network has been widely used in the field of computer vision and achieves remarkable results.The lung cancer has been a serious threat for human health,hence,it's very important to make diagnosis and treatment early for lung cancer patients based on digital images.Thus,computer aided diagnosis as the doctors' second eyes can provide accurate quantitative analysis and make up for eyes' inertia and gray-scale insensitivity defects,it can help doctors diagnose quickly and improve the consistency of the diagnosis results.Objectives CT?PET and PET/CT images of lung tumor were regarded as the research object,a novel convolutional neural network was proposed to realize the computer aided diagnosis of lung tumor,the purpose is to help doctors achieve accurate diagnosis,reduce labor intensity,promote the computer intelligent process and realize the deep learning application in the medical field.Methods Based on the model structure of convolutional neural network,the ensemble convolutional neural network and the deep convolutional neural network were proposed for lung tumor recognition.Firstly,three convolutional neural networks(CT-CNN,PET-CNN,PET/CT-CNN)were constructed in different sample spaces,the ensemble convolutional neural network was used to recognize lung tumor based on local feature.Secondly,the deep convolutional neural network is used to recognize lung tumor based on global features.The recognition rate,operation time,sensitivity,specificity,MCC and F1 Score six indexes were used to evaluate convolutional neural network recognition performance.Results For the ensemble convolutional neural network,three comparative experiments were done,the recognition performance evaluation of single convolutional neural network,the influence of different model parameters on recognition results,comparison experiments among ensemble convolutional neural networks,a single CNN and other traditional methods.The experiment results show that the convolutional neural network is feasible for computer aided diagnosis of lung tumor,the iterations and the batchsize have an effect on the identification results,and the ensemble neural network is better than the single CNN and the traditional identification methods,it's proved the superiority of the algorithm.Three comparative experiments were performed on the deep convolutional neural network,comparative experiment with different model parameters,comparative experiment with different structure models and comparative experiment with different optimization methods.The experiment results show that the deep convolutional neural network is feasible to recognize lung tumor based on global features,and it can achieve a good recognition rate when the appropriate model parameters and model structures were selected,the method of gradient descent with momentum was adopted.