Research Of Lung Nodules Segmentation And Classification Method Based On Thin CT Image Sequences

Lung cancer is the malignant tumor with the highest morbidity and mortality among cancers worldwide.With the rise of artificial intelligence and the application of deep mining of large data combined with optimization of computational model,computer-aided diagnosis has greatly improved the automatic diagnosis rate of lung cancer.However,because the clinical diagnosis of lung cancer is mostly mid-late(stages II,III,and IV),the 5-year survival rate is still less than 5% after standardized treatment(surgery or stereotactic radiotherapy),and the cure rate of early lung cancer patients can reach More than 80%.However,early features of small lesion are difficult to detect,and CT imaging under thin-slice scanning mode can detect early lung micro-lesions sensitively.At the same time,CT imaging technology brings a large number of CT image sequences,which will inevitably lead to the difficulty of CT image processing in the computer aided diagnosis system,slow processing speed and low efficiency.Accurate segmentation of nodules in CT sequence image and improving the accuracy of classification of benign and malignant lung nodules will greatly improve the efficiency of computer-aided diagnosis system.Therefore,segmentation and classification have become the focus and difficulty computer-aided diagnosis of lung.In this paper,the methods of segmentation and classification of lung nodules in thin-scan CT images are discussed and studied.(1)In order to improve accuracy and efficiency of segmentation of ground glass opacity nodule,a Sequence segmentation for Ground Glass Opacity Nodule based on automatic seed point was proposed.Firstly,adaptive Gauss filtering is applied to segmented lung parenchyma images to stretch the contrast between lung parenchyma,ground glass nodules and blood vessels.Then,the improved sliding algorithm was used to locate the seed points automatically and accurately.Finally,the image of lung nodules was obtained by iterative fuzzy connectivity algorithm.The experimental results show that this method not only has a good segmentation effect for ground glass lung nodules,but also can achieve accurate segmentation for isolated and pleural traction lung nodules.Moreover,the proposed algorithm solves the problem of artificial interaction between seed points,and improves the segmentation effect of lung nodule image as well as the generality of the algorithm.(2)In the classification of lung nodules,a classification method of lung nodules based on multi-scale spatial pyramid pooling was proposed.Due to the different size of lung nodules,when the same scale image is input into the convolutional neural network,there will always be more background redundant information,and the lung nodule features cannot be accurately extracted.In order to solve this problem,Firstly,the classic AlexNet network was improved to make it more suitable for the classification of images of lung nodules.Secondly,a multi-scale input was designed to reduce the redundant information of ROI.Finally,the image features of lung nodules were extracted by Multi-Scale Spatial Pyramid strategy and a fixed length representation was achieved.The experimental results show that compared with other methods,this method has better performance in accuracy,sensitivity and specificity.The method can achieve higher classification accuracy.