Research On Detection Method Of Pulmonary Nodules Based On Ensemble Convolutional Neural Network

Lung cancer is the most common malignant tumor in the world and the leading cause of cancer death.Early detection of lung cancer is highly difficult.If lung cancer patients are diagnosed at the early stage and undergo relevant treatment,the 5-year survival rate will be greatly improved.Reading a large number of patients with lung CT images is a high-intensity and time-consuming task for doctors.The doctor's fatigue during reading is enhanced,which leads to a decrease in the efficiency and quality of the reading images,resulting in misdiagnosis and missed diagnosis.Occurrence,which is one of the important reasons for lung cancer mortality in recent years.The computer-aided diagnosis system can classify and detect the CT images of the lungs through relevant algorithms,which can reduce the workload of doctors and provide a powerful guarantee for the early prevention and monitoring of lung cancer.Traditional detection methods have the disadvantages of accurately segmenting lung nodules and manually extracting image features.The use of convolutional neural networks for lung nodule detection has begun to spread,but the current detection models are mostly single-input single-channel network structures,no matter the network structure.The depth of the complex layer can only be extracted for the lung nodules of a certain scale,but the reality is that the lung nodules themselves are the same size,and the non-pulmonary nodules as negative samples are also not uniform in scale,which gives A single-channel network poses a great challenge,which in turn affects the further improvement of detection accuracy.Moreover,some networks currently have very complicated structures in order to mine more information.In actual operation,they will encounter computational bottlenecks and over-fitting phenomena,which will also affect the detection efficiency while occupying a large amount of resources.In view of the above drawbacks,this paper focuses on the multi-scale input integrated convolutional neural network lung nodule detection method.The specific work includes the following aspects:(1)Proposes a method for accurately extracting regions of interest in the lung based on XML annotation files.In the LIDC-IDRI(The Lung database Consortium and Image database resource Initiative)public database,the lung nodules and non-pulmonary nodules are accurately extracted using the annotation information of experts in the XML file,and constructed by normalization processing and data enhancement.A sample library of positive and negative sample balance.(2)The traditional machine learning algorithm is used to identify the disadvantages of lung nodules,and the deep learning idea is introduced.The convolutional neural network model is established for the classification and identification of pulmonary nodules.The Dropout ratio,the number of fully connected layer feature maps,and the activation function are determined through experiments.Selecting important network parameters,the biggest advantage of this algorithm is that the model can automatically extract the deep features of the image,and the classification result is more objective and real.(3)Based on the problem of insufficient performance caused by single-input single-channel in convolutional neural network,an ensemble learning idea is introduced and an integrated convolutional neural network model based on multi-scale input is extracted.Through multi-scale image input,three weakly classifiers are respectively determined by using three different networks.Finally,the results of the weak classifier are fused by a voter mechanism to obtain a strong classifier.Such a network structure is more than a single scale input.The picture can learn more characteristic information of lung nodules,thereby improving detection performance.In this paper,experiments on the LIDC-IDRI database showed that the recognition accuracy,sensitivity and specificity reached 93.0%,93.3%,and 92.7%,respectively,which had certain advantages compared with the current latest lung nodule classification algorithm.In this paper,the convolutional neural network model,multi-scale input concept and integrated learning thought are introduced into the lung nodule detection and recognition experiment.The results show that the proposed algorithm can help doctors to diagnose pulmonary nodules to a certain extent,and is an early discovery of lung cancer.Early intervention,early treatment "provides technical support.