Research On Key Technology Of Lung Nodule Detection Based On Deep Learning

Lung cancer has become the first cause of tumor death in China.Clinical diagnosis shows that the five-year survival rate of patients with lung cancer depends largely on whether it is early or late when it is detected.The earlier the patient with lung cancer is detected and treated,the higher the five-year survival rate.However,lung cancer is usually asymptomatic in the early stages,and more than 80% of lung cancer patients are found to be advanced when they are found,thus losing the best opportunity for treatment.Therefore,early diagnosis of lung cancer can not only improve the treatment effect,but also improve the survival rate of patients.Thoracic CT images are the most basic basis for medical personnel to detect and diagnose lung cancer intuitively through human visual observation.However,with the increasing resolution of CT imaging,the volume of lung nodules that can be detected is getting smaller and smaller,which can provide more tissue information,and also bring a great workload for physicians,which causes physician fatigue Distraction,increase the probability of misjudgment and missed judgment.In recent years,the rapid development of deep learning has brought a new idea to lung nodule research.The existing methods have problems such as insufficient accuracy and high false positive rates.Aiming at the above problems,this paper proposes an improved lung nodule detection algorithm based on Faster R-CNN and a model-based multidimensional false positive reduction algorithm.The specific work is as follows:(1)To improve the accuracy of existing lung nodule detection algorithms,improve Faster R-CNN.First,replace the VGG16 network in the original model with a residual network,and add an inception module to optimize feature extraction.The internet.Next,a multi-feature fusion feature pyramid structure FPN(Feature Pyramid Network,FPN)is introduced to fuse the location information of the underlying feature map and the semantic information of the high-level feature map to obtain more semantically rich features with both location information and semantic information.Figure,and then modify the anchor size in the RPN network,so that nodule candidate box generation is more efficient.On the public data set LIDC-IDRI,the method achieves detection accuracy of 89%,respectively.(2)Aiming at the problem of high false positive rate in lung CT image nodule detection tasks,this paper proposes a model-based multidimensional false positive reduction algorithm.Specifically,a two-dimensional binary classification model and a three-dimensional binary classification model are designed.Model integration of these three models can make full use of the nodule space context information.Non-nodular 3D images clearly show non-spherical linear structures,which allows us to distinguish nodules from false positives.The accuracy of the model on the public data set LUNA16 is 93.81%.The experimental results show that the multi-dimensional vacation positive removal algorithm based on model integration can effectively reduce false positives and improve the accuracy of the model.