Research Of Pulmonary Airway And Vascular Segmentation Based On Self-Supervised Transfer Learning

The segmentation of pulmonary vascular and airway in CT images have significant value in research for the diagnosis-treatment of lung diseases.Because traditional segmentation methods are based on edge detection,region growth,or specific models for image processing,a large amount of human-computer interaction must be used to improve the segmentation accuracy of the dense blood vessels or airway trees in the lungs.In recent years,deep neural networks have gradually been applied in the pulmonary airway and vascular segmentation to ensure automation,accuracy,and efficiency.However,its effect is severely affected by the quality and scale of the annotation data,which limits the promotion of such methods.To solve this problem,a new method for pulmonary airway and vascular segmentation based on self-supervised transfer learning is proposed in this paper.The main research contents of this thesis are as follows:1.Lung feature extraction for airway and vascular segmentation based on self-supervised learning.Since the deep learning model accuracy is severely limited by the labeled data,this paper uses the 3D anatomical information of the unlabeled CT image itself to obtain general features of the lungs through self-supervised learning,and then transfer to the segmentation task of lung trachea and blood vessel.Focus on the characteristics of the lung trachea and vessels such as complex and diverse structures,multi-scale and high noise,the image restoration pretext task is used for self-supervised learning,and five image transformation methods are used to construct pseudo-labels.In this way,the model is guided to comprehensively learn lung features to meet the needs of different levels of characteristics for the segmentation task of airway and blood vessels.2.Pulmonary airway segmentation based on the fine-tuned self-supervised model.In the process of the fine-tuning self-supervised model with a small amount of lung tracheal annotation data,the block-by-release strategy is applied to determine the optimal depth of network fine-tuning.To overcome the imbalance of foreground and background voxels in the tracheal annotation image,specific data preprocessing methods and the fine-tuning strategy incorporating supervision information are adopted.Meanwhile,the method of data enhancement is used to improve the quality of training data,thereby improving the accuracy of lung vessel segmentation.3.Pulmonary vascular segmentation based on feature extraction using the self-supervised model.For the lung vessel segmentation task,a small number of discrete annotation points is used as supervisory information.To solve the problem that the end-to-end network training must use the full image annotation,the strategy of feature extraction and classification training in stages is adopted.The self-supervised pre-trained model is used as the feature extractor,and the best model combination for pulmonary vascular segmentation is studied.Because a few sparse vascular labeled points cannot cover the whole vascular tree,it is not enough to evaluate the performance of the segmentation algorithm.Therefore,a test dataset suitable for this task is constructed.Experimental results show that the proposed algorithm can achieve high-precision automatic segmentation of pulmonary airway and vascular in CT images with only a small number of labeled samples,and have better performance in terms of time,accuracy,and robustness.