Lung Nodule Detection And Lung Cancer Recognition Based On 3D Convolutional Neural Network

Lung cancer is one of the leading causes of cancer deaths in the world,among both women and man,and it is estimated that 1.6 million people die of the disease each year.For lung cancer patients,the sooner the lung cancer is found,the higher its survival rate is.Currently,Radiologists usually need computerized tomography scans of patients to examine abnormal areas of lungs to diagnose lung cancer.The recent advancements of deep learning methods have made possible for achieving several human-level tasks such as image classification,voice generation and game playing.They were also successfully applied on medical applications for dermatologists-level skin cancer detection and brain tumor segmentation.Apart from traditional computer vision tasks,lung cancer detection is enormously hard mainly for(1)Many nodules are labeled by non medical experts which will easily cause a large portion of false positives.(2)CT images scans are in three-dimensional volumes will make any detection algorithms computationally expensive.(3)The unbalancing data lack of diversity and for validation.Aiming at the CT scan images of the lungs,this paper studies a lung cancer detection and recognition algorithm based on 3D convolution neural network to assist doctors in the diagnosis of lung cancer.The algorithm is mainly divided into three parts,image preprocessing,lung nodule detection and lung cancer identification.In the image preprocessing,we mainly carry out the following steps: converting the gray value of the CT image into the HU value,and then reusing the pixel for the purpose of unifying the corresponding size of the pixel in the physical world,followed by extracting Lung mask to obtain the CT image of the lung cavity area for better processing,and finally the standardization of data processing in order to better adapt to the network input requirements.In the detection of pulmonary nodules,divided into three part.First,regional block division is used to solve the problem of GPU memory shortage.Then the positive and negative sample balance is used in order to solve the imbalance of positive and negative samples in the training data.Last,building Lung nodule detection network for detecting candidate nodules.In the lung cancer identification section,candidate lung nodules were screened as input of the lung cancer recognition network,which outputs the probability of lung cancer in the patient.The algorithm detects suspicious nodules and gives the probability of patients suffering from lung cancer.The former can identify suspected nodules for doctors’ reference,and the latter can give a diagnosis of lung cancer probability.Both of them can help doctors to diagnose lung cancer.In this paper,a detailed experiment is carried out for the proposed algorithm.The experimental data consists of two parts,one part is open source Kaggle dataset,a total of 2285 images in the dataset,including 2359 pulmonary nodules,of which the diameter of 1186 nodules is less than 35 mm and 905 nodules’ diameter is less than 10 mm Nodules 905;the other part is a multi-center dataset from a number of hospitals in China,a total of 820 samples,of which435 cases is benign samples,385 cases is malignant samples.Our algorithm model has high specificity and sensitivity for the classification of lung cancer by validating experiments on the above dataset.