| Lung cancer has the highest incidence all over the world.Unfortunately,people have not yet found an effective method for reducing the incidence.Moreover,the patients with advanced lung cancer can hardly be cured under existing medical conditions.Therefore,early detection and treatment is an important way to reduce the mortality of lung cancer.Early-stage lung cancer generally manifests as pulmonary nodules.Chest CT is currently used to screen for pulmonary nodules.There are hundreds or even thousands of CT images for every patient.If a doctor personally views the CT images for screening,two problems will arise.First,the huge workload can easily cause doctor fatigue and increase the probability of misjudgment.Second,manual screening relies too much on the doctor’s clinical experience.It is also susceptible to the subjective judgment of the doctor.Therefore,the use of computers to screen for pulmonary nodules has significant clinical significance.In recent years,pulmonary nodule detection methods based on deep learning have been widely used.The output features can be directly used for the identification and classification of pulmonary nodules.The methods have a high degree of automation and a fast detection speed.In addition,the methods can effectively distinguish real nodules from non-nodules.However,there are many types of pulmonary nodules with different sizes and varying positions.How to detect pulmonary nodules accurately and efficiently is still worthy of attention.In addition,the difficulty of detecting small pulmonary nodules is another important point.In response to the above problems,the thesis proposes a pulmonary nodule detection method based on convolutional neural network.The main work in the thesis is as follows:(1)The proposal of pulmonary nodule detection algorithm.Based on reading domestic and foreign papers and investigating the research status in this field,the thesis proposes a one-stage pulmonary nodule detection algorithm without any false positive reduction.All calculations are encapsulated in a network.(2)The design of network.The thesis uses 3D ResNet as the backbone network,which can not only improve the model’s ability to extract the spatial features of medical data,but also avoid gradient disappearance and gradient explosion during deep training.A multi-scale module is designed to optimize the semantic features.The module uses a combination of multiple convolutions to obtain different sizes of receptive fields,so that the network can extract the features of pulmonary nodules at different scales.For small pulmonary nodules,an attention module is designed to mine the correlation information between features from the perspectives of space and channel and strengthen feature transmission and reuse.Important features are highlighted,and unnecessary features are suppressed.Then,the enhanced features are input into a pyramid fusion module,so that the features contain both deep texture information and shallow position information to obtain better positioning and classification.In order to solve the problem of imbalance between the samples in training,the thesis introduces Focal Loss function to adjust the weight of the samples.(3)Experimental verification.Experiments are performed on the LUNA16 data set.Compared with other methods,the method proposed in the thesis obtains better detection results.Automatic detection of pulmonary nodules is of great significance for early lung cancer screening.It can buy precious treatment time for patients,thereby greatly improving the survival rate of patients. |
