Detection Of Low Dose CT Pulmonary Nodule Based On Improved Faster R-CNN

Lung cancer is a malignant tumor with the highest morbidity and mortality.As the haze situation has intensified in recent years,the number of people suffering from lung cancer has increased year by year,and society has paid more and more attention to this.Studies have found that effective treatment in the early stages of lung cancer can improve survival by 65%.Lung cancer patients have no obvious symptoms in the early stage,and the early stage of lung cancer is reflected by the form of lung nodules.Therefore,the early diagnosis of lung cancer is the screening of lung nodules.Therefore,effective screening of lung nodules is of great significance.This article first created a low-dose CT lung nodule dataset,taking low-dose CT lung nodules as the research object,and using the deep learning method Faster R-CNN algorithm as the technical means,aimed at providing doctors with more efficient auxiliary diagnosis Plan,laying the foundation for subsequent effective treatment.Currently,there are three problems in lung nodule detection based on Faster R-CNN.First,the feature extraction network is weak in extracting nodule features;second,the regional recommendation network anchor box does not match the lung nodule data set;Third,the method of selecting difficult cases for classification regression network is not representative.In view of the above problems,this paper proposes an improved algorithm based on Faster R-CNN,which solves these three problems well.The main work of the thesis is as follows:Construct a low-dose CT lung nodule dataset suitable for deep learning.Firstly,CT images with nodules were selected based on the coordinate information of the LIDC dataset,and the CT images were transformed into a PNG format that can be input into a convolutional neural network.Second,nodule coordinate information was used to label the images with nodules.The data is amplified by flipping to form a lung nodule data set that can be input into the target detection frame.Aiming at the problem that the feature extraction network has insufficient ability to extract nodule features,this paper first selects Resnet101 to replace VGG16 as the feature extraction network of Faster R-CNN,followed by enhancing the learning ability of the feature network,introducing the channel attention mechanism SENet,and based on it A second-order response transformation strategy is added to build an improved feature extraction network.After this step of improvement,the feature extraction capability of the Faster R-CNN feature extraction network is enhanced,and the lung nodule detection accuracy is improved by2.92%.Aiming at the problem that the regional recommendation network anchor box does not match the lung nodule dataset,this paper uses K-Means ++ clustering method to complete the prior information analysis of the lung nodule dataset,and reset the anchor box size and proportion based on the clustering results.Achieve improvements to regional recommendation networks.After this step improvement,the problem of anchor frame setting and data set mismatch is solved,and the detection time is shortened on the basis of ensuring the accuracy of lung nodule detection.Aiming at the problem that the selection method of classification regression network is random selection,this article introduces an online hard example mining algorithm(Online Hard Example Mining,OHEM)into the classification regression network,and selects the negative samples through the online loss ranking method to select the difficult cases.,And then let the network learn the hard cases that have been screened to improve the classification and regression network.After this step improvement,the classification regression network's ability to detect difficult negative samples is improved,and the accuracy of lung nodule detection is improved.