Deep Learning For Automated Segmentation Of Pulmonary Structures

Because of air pollution,smoking and bad living habits,more and more people are suffering from lung cancer,pneumonia and other lung diseases.Doctors need to analyze the lung image to diagnose disease and make operation plan.At present,doctors can only adjust window width and position to deal this problem.Therefore,it is of great research value to segment lung organs by computer image processing and improve the efficiency of doctors’ diagnosis.The traditional method of lung organ segmentation is tedious and low automation,while using deep learning to realize pulmonary structures segmentation can realize end-to-end segmentation,simplify steps and reduce the time required for segmentation.But there are still many shortcomings.First of all,the pulmonary structures segmentation task lacks annotation data for network training.Secondly,at present,there is few research on data set construction methods for medical images.Most medical images are three-dimensional images.Using different data set construction methods will affect the segmentation accuracy.Finally,the u-net network,which is widely used in the field of medical image segmentation,still has room for improvement in segmentation accuracy and network model size.Therefore,this paper studies the above problems and proposes solutions.First of all,based on the characteristics of lung medical image,this paper uses traditional methods to segment three lung organs from CT image:lung parenchyma,lung trachea and lung vessels.According to the shape and characteristics of the segmented object,we propose adjacent slice optimization,ignoring edge filtering,ROI cycle growth and other methods to optimize the segmentation results,so as to improve the integrity of segmented organs.Secondly,three different data set construction methods and five different data augmentation methods are proposed to analyze the influence of different methods on the training loss and segmentation accuracy of neural network,and to find out the most suitable data set construction and data augmentation methods for network training.Finally,according to the characteristics of lung medical image,two network structure optimization methods based on u-net network are proposed,which can reduce the parameters of network model and improve the accuracy of network segmentation.