Automatic Detection Of Lung Nodules And Lung Segment Evaluation Based On Deep Learning

In recent years,due to environmental and other factors,the number of lung cancer patients has increased year by year.Lung cancer has become one of the most important cancers in China and many countries around the world.Lung cancer is a malignant tumor with a high mortality rate.The 5-year survival rate of early patients is about 49%,and the 5-year survival rate of advanced patients is less than 1%.Therefore,the screening and treatment of early lung cancer has become a mortality way to reduce lung cancer.Important means.Early lung cancer manifests in CT images as small nodules with a diameter of 1mm-5cm,and the CT data of a patient’s lung generally contains 300-500 pictures,which is not easy to be discovered,which brings certain difficulties to the diagnosis of early lung cancer.Therefore,how to apply computer image processing technology to early automatic screening and surgical evaluation of lung cancer is of great significance.The specific work content of this article is as follows:(1)Propose a method of combining deep learning methods with traditional methods to segment the lung parenchyma to provide a reference for the subsequent detection of lung nodules and segmentation of lung segments to improve the accuracy of the final result.After segmenting the lung parenchyma using 2D Unet,the ABM(Adaptive border marching)method was used to process the subpleural nodules.Five sets of CT data containing near-pleural nodules were randomly selected from the Luna16 data set for testing.The results showed that The near pleural nodules in the five sets of data were correctly segmented,and the final lung parenchymal segmentation DICE was 98.9%.(2)Aiming at the existing target detection framework,an improved 3D Unet network structure is proposed,which is easier to train than the two-stage target detection framework and has a sensitivity of 86.9%in lung nodule detection tasks.In addition,DF Net(Dual Path Feature pyramid Network)is proposed for the characteristics of small targets and wide diameter distribution in the detection of lung nodules,and the final detection sensitivity of lung nodules reaches 88.2%.(3)For the evaluation of early lung cancer surgery,this paper segmented the lung segment,first proposed an algorithm based on memory area growth to segment the lung trachea,then extracted the center line of the lung trachea,and manually selected the branch point of lung segmentation The lung segmentation model was generated according to the tracheal direction,and the results of the algorithm for segmenting lung segments were compared with the results of manual labeling.DICE reached 83.3%.