CT-based Computer Aided Diagnosis System For Lung Cancer

Among various common malignant tumors,the number of new patients and new deaths caused by lung cancer account for the largest number of various cancer-causing deaths.Although the lung cancer mortality rate dropped by 17% from 2002 to 2014 and by 43% from1990 to 2014,the global number of deaths was 1.59 million,accounting for 26% of all cancer deaths.An important basis for the early diagnosis of lung cancer is CT medical imaging.Additionally,it is necessary to use CT to guide the puncture area and mark the contour of the lung nodules according to CT,along with the subsequent diagnosis and treatment plan.Therefore,if the computer-related technology can be used for preliminary identification diagnosis and contour marking of the lung nodules,it will be of great help to the clinical diagnosis.To assist doctors in the clinical diagnosis of lung cancer nodules,we build a CT-based computer aided diagnosis system for lung cancer.In order to achieve the goal of lung nodule recognition,we propose a pre-processing process for CT images,including CT visualization,lung region segmentation,and nodule ROI construction.We used different types of networks to test on the proprietary lung nodule data set provided by Wuhan Union Hospital,and finally the modified VGG16 is chosen as the classification model.Besides,Experiments show that the proposed pre-processing process can improve the accuracy of the network model by about3% compared with the method of only using a single slice.Another important task is to interpret the output of the neural network into diagnostic information that can be understood by doctors.The classification value output by the neural network is not conducive to doctors,and carries little information.To solve this problem,we uses the Grad-CAM algorithm to extract the gradient update value of the neural network after processing the input,and uses it to calculate the heat map representing the region of interest of the neural network classification.The color depth of the heat map represents the degree of confidence of the neural network to confirm the input sample as a malignant nodule.Moreover,the area covered by the heat map also covers the location of the lung nodules,so it can be good for the doctor’s diagnosis and subsequent puncture examination.Experiments show that,in the case of the heat threshold chosen as 0.6,our generated heat map achieve the best coverage effect,the Dice coefficient is 67.87%.On the basis of CT processing,CNN classification and interpretation,combined with a number of auxiliary functions,such as CT windowing,our diagnosis system automates the process of doctors’ diagnosis of lung cancer to a certain extent,improving the efficiency of clinical diagnosis.