Research On Blood Vessel Segmentation In Pulmonary CT Images Based On Deep Learning

Lung disease is an important area of medical research.Both of the outbreaks of COVID-19 all over the world in 2020 and SARS in 2003 are the diseases caused by virus infection in the lungs.At present,research on various lung diseases is focusing on the changes of pulmonary blood vessels,such as pulmonary arterial hypertension,vascular lesions,arteriovenous malformations,etc.In order to detect and treat diseases in advance,Computer Aided Diagnosis(CAD)is usually utilized.Rapidly and precisely segmenting pulmonary blood vessels and finding out the region of interest are both the important steps for CAD,so the pulmonary blood vessel segmentation has attracted widespread attention of researchers.In CAD,Computed Tomography(CT)is widely used for its advantages such as fast scanning,less damage to the body,a large amount of raw data,and reconstruction at different layer distances etc.For lung diseases,at present the further diagnosis is mainly based on lung shadows provided by CT.However,the CT image has a large amount of data,and the traditional manual blood vessel segmentation based on CT images is time and labor consuming with low precision.Therefore,to explore automated pulmonary blood vessel segmentation is helpful to improve the efficiency and accuracy of lung disease diagnosis.Based on the original pulmonary CT images,this thesis focuses on the segmentation of the pulmonary mediastinum window and pulmonary blood vessels.The main research includes the follows:1.About 4789 original CT images of 6 patients from a hospital in Gansu Province were labeled twice,resulting in a dataset: the first was to mark the two crescent-like pulmonary mediastinum windows,and the second was to mark the pulmonary blood vessels.2.This thesis designed a layered model based on original CT images,produced separate training and test sets for each layer,and then combined BCDU-Net based on deep learning for edge detection and segmentation of pulmonary mediastinum window.3.After segmenting the pulmonary mediastinum window,a pulmonary blood vessel segmentation model is designed,and then Support Vector Machine,Gaussian Pyramid and Sparse Autoencoder were combined together to realize the automated blood vessel segmentation.The experimental results show that the method proposed in this thesis has achieved a good segmentation effect,and has certain reference value and promotion for the further application of CAD in the study of lung diseases,and it can also be used as the work basis for future related research.