Design And Implementation Of Aided Detection System For Pulmonary Nodules Based On Convolution Neural Network

Lung cancer is one of the most common malignant tumors in China,and its mortality ranks first among cancer deaths.Since the vast majority of patients are diagnosed with advanced lung cancer,the 5-year survival rate of patients with lung cancer is only 16.1%.As an early manifestation of lung cancer,pulmonary nodules can be diagnosed by lung CT images,which plays a key role in the diagnosis of lung cancer.With the rapid development of computer level and artificial intelligence,the use of deep learning algorithm to detect CT images has become the main research direction in the field of lung nodule detection.This method of using computer to detect lung nodules can greatly reduce the workload of doctors and help doctors diagnose lung cancer accurately.Starting with CT images,this paper studies the detection methods of pulmonary nodules.By studying the literature,the current research status and methods in the field of pulmonary nodule detection are collated and analyzed,and the methods based on manual extraction and automatic extraction are compared.In order to solve the problem of low accuracy of the above two methods,a simple full convolution model YOLACT for case segmentation is introduced.Both YOLACT model and Faster R-CNN belong to the target detection network,and the detection accuracy and speed in the field of target detection are better than Faster R-CNN.In order to verify that the model can effectively detect pulmonary nodules,the false positive screening experiment was carried out at first,and the detection results of the model were evaluated in all directions.The false positive rate and missed diagnosis rate of the model were tested while using the general evaluation index,and the experimental results were analyzed.It is proved that YOLACT model can effectively detect pulmonary nodules,but the detection accuracy and other indicators still have room for improvement.In this paper,an improved YOLACT model is proposed:In the backbone network of the model,DetNet is used to replace the original ResNet,to enhance the ability of local feature extraction and improve the accuracy of detection,and in model training,transfer learning mechanism is introduced to accelerate model training and reduce the amount of computation.In the activation function,the RReLU activation function is used to replace the original ReLU activation function to solve the situation that some weights can not be updated,and improve the accuracy of lung nodule detection.The detection accuracy of the improved model on the LUNA 16 data set is 94.58%,which is higher than that of the mainstream lung nodule detection algorithm.Finally,using the improved YOLACT model,an auxiliary detection system of lung nodules is designed and implemented to realize the automatic detection of lung CT images,so as to reduce the workload of doctors and improve the efficiency of lung cancer diagnosis.