A Lung Cancer Detection Model Based On Deep Learning

Lung cancer is the second largest cancer,and about 14% of cancer patients in the world are lung cancer.The reason that most patients are not cured is that they have reached the advanced stage of the disease when lung cancer is found.If the patient can be detected in the early detection of lung cancer,the patient's survival rate also increased significantly.Therefore,the detection of lung cancer began to gradually receive attention.At present,CT(computed tomography)scanning can be performed on the patient's lungs to obtain CT slices.The doctor determines whether the node is benign or malignant by the morphology of the nodes in the CT slice to determine whether there is lung cancer.However,this process is entirely dependent on the experience of the doctor.Different doctors may give different results on the same CT slice,and in some cases,missed detection may occur,which may adversely affect the treatment of lung cancer patients.In the field of computer science,how to deal with the CT scan of the lungs by establishing a model and then carry out automatic lung cancer detection and to analyze these nodes to determine the probability of the patient's illness has become an important topic in computer vision field.Different from the two-dimensional image classification,segmentation,target recognition and other tasks,The following challenges exist in the detection of lung cancer: Firstly,the amount of medical image data is very small compared with two-dimensional image classification.For example,there are 1 million training data for the image classification task of the Large Scale Visual Recognition Challenge(ILSVRC).For LIDC-IDRI,there are only 1018 samples,with relatively little data.In this paper,a lung cancer detection system suitable for three-dimensional CT images is proposed.The detection classification model of lung cancer nodes is designed to improve the accuracy of lung cancer detection.The research contents of this paper are as follows:(1)Summarize traditional and in-depth lung cancer classification and briefly describe the pros and cons of previous research methods.(2)In order to solve the problem of fewer detection datasets of lung cancer,this paper presents data processing and enhancement methods suitable for the location of lung cancer nodes.(3)In order to solve the problem of two-dimensional image models which are not suitable for three-dimensional image,this paper presents a new three-dimensional convolution neural network model.The model consists of two parts.The first part is a three-dimensional fully convolutional network(FCN)model that generates a heat map of lung cancer nodes.From the heat map,we can locate the location of those malignant nodes.According to the heat map generated by the first part,the second part selects those vicious nodes with high probability and then merges the features of these selected nodes into a feature vector,which represents the whole situation of lung scanning.Finally,we use this feature to classify to determine whether it is suffering from lung cancer.(4)Finally,we carried out experiments on the public dataset LIDC-IDRI to prove the effectiveness of the proposed method.