Research On Detection Of Pulmonary Nodules Based On Convolution Neural Network

Lung cancer has the highest mortality rate among cancers,and early diagnosis is crucial for patients.Pulmonary nodules are one of the early manifestations of lung cancer,and the most commonly used detection method is lung CT images.Doctors need to spend a lot of time reviewing the patient’s CT images,which can easily cause doctors to experience fatigue in reading images,leading to misdiagnosis or missed diagnosis.By using convolutional neural networks to assist doctors in detecting pulmonary nodules,not only can the workload of doctors be greatly reduced,but also machine diagnosis has a certain degree of objectivity and the detection results are more stable.This thesis adopts a two-stage pulmonary nodule detection method.In the candidate nodule detection stage,it is mainly necessary to ensure a high sensitivity and tolerate a certain degree of false positives.To reduce the false positive stage,it is necessary to reduce the presence of false positives as much as possible,so that the overall pulmonary nodule detection model has both high sensitivity and low false positives.The specific research content is as follows:(1)According to the particularity of lung CT images,data preprocessing is performed in the steps of lung parenchyma segmentation,normalization,and data enhancement.In the segmentation of lung parenchyma,this paper uses the global threshold method for threshold segmentation.In data enhancement,data volume is expanded through data enhancement,and the problem of unbalanced positive and negative samples is solved.(2)Aiming at the problem that most pulmonary nodules have small diameters and are prone to miss detection,a candidate nodule detection model based on an improved U-Net network is proposed.This model takes U-Net network as the basic skeleton,and adds a residual module in the encoder part,which improves the problem of insufficient depth of U-Net network.Replacing jump connections with a hybrid attention mechanism module makes the model more focused on learning useful information,effectively transmitting shallow feature information into the deep layer,making semantic information and feature information fusion more sufficient,and alleviating the problem of small target detection.In the decoder section,a deep supervision module is added to provide relay supervision for the network.At the same time,the feature results output from the decoding layer are combined as the final output of the network,making full use of hidden layer features.Experiments show that the sensitivity of the candidate nodule detection model on the LUNA 16 dataset reaches 96.7%.(3)To address the problem of high false positives in candidate nodule detection model,a multi-dimensional integrated false positives reduction model based on dense connections is designed.The model includes three submodels,namely,a twodimensional binary classification model,a three-dimensional binary classification model A,and a three-dimensional binary classification model B.All three submodels are composed of dense connection modules.In order to solve the problem of large differences in the size and shape of pulmonary nodules,this false positive reduction model uses three different scale images as input to the three submodels,and the final classification result is obtained by adding the weighted average of the three submodels.This model can not only extract different levels of contextual information using information from different receptive fields,but also integrate two-dimensional and three-dimensional information from lung CT images.The experiment shows that the CPM of this false positive reduction model reaches 0.915 on the LUNA16 dataset.The CPM of the overall pulmonary nodule detection model reaches 0.876,the average number of false positives per scan is 15.6.