Research On Computer-aided Diagnosis Of Lung Cancer From CT Scans Based On Improved U-Net

Lung cancer is one of the malignant tumors with high mortality among all kinds of malignant tumors.There are many deaths due to lung cancer,and the five-year survival rate is very low.The early symptoms of lung cancer are not obvious,and patients are easy to miss timely treatment.Therefore,regular and early screening is helpful to the prevention and treatment of lung cancer and improves the survival rate of patients.Low dose CT has become the first choice for early screening of lung cancer because of its advantages of high sensitivity and low radiation.One of the early forms of lung cancer on CT scan is pulmonary nodules.Radiologists can diagnose patients according to the characteristics of nodules.Therefore,it is very important to accurately detect pulmonary nodules and give the malignant probability for the diagnosis of lung cancer.The computer-aided diagnosis system of lung cancer can assist radiologists to detect nodules on CT images and give benign and malignant information,which can effectively reduce the working pressure of doctors.At present,the deep learning algorithm which can automatically learn the feature of nodule images is gradually taking the mainstream in the field of lung cancer assisted diagnosis than traditional machine learning algorithms.This thesis summarized and studied the existing pulmonary cancer aided diagnosis technology using deep learning.On this foundation,a threedimensional convolutional neural network was designed for feature extraction of lung nodule detection and diagnosis.Firstly,a 3D U-Net was selected as the basic framework of the detection network,and multiple 3D Res Ne Xt modules were added to the U-Net to improve the accuracy while maintaining the complexity of the model by adding multiple groups of convolution.Then,two attention modules which can automatically learn the characteristics of nodules with different sizes and shapes were added to the skip connection of U-Net to form an improved U-Net.Then,the constructed improved U-Net was used for nodule detection,and the candidate nodules were directly output through the RPN output layer.In particular,the generalized intersection over union was selected to calculate the confidence and label when outputting the candidate nodules,so as to better optimize the network parameters.Finally,in order to avoid the over fitting problem caused by limited data,the candidate nodules generated by detection were input into this improved U-Net again to extract benign and malignant features,and the leaky noisy or method was used to integrate the cancer probability of multiple nodule candidates in each CT scan.The diagnosis model of lung cancer based on improved U-Net proposed in this thesis is evaluated on DSB dataset,and the ideal experimental results are achieved.The sensitivity,specificity,accuracy and G-mean of pulmonary nodule detection are 98.15%,99.99%,99.99%and 99.07,respectively.The accuracy of lung cancer diagnosis reached 80.43%,AUC value reached 0.86,cross entropy loss and mean square error decreased to 0.43 and 0.14,respectively.To sum up,the lung cancer diagnosis framework based on improved U-Net proposed in this thesis is evaluated on DSB dataset,and the experiments are designed in this thesis has advantages and satisfactory results.It proves that this framework is highly sensitive to the characteristics of pulmonary nodules.Moreover,this model can enhance the performance and reduce the loss of pulmonary nodule detection and diagnosis,which has a certain clinical value.