Automatic Pulmonary Segmentation Algorithms For Tuberculosis CT Image

Tuberculosis disease is a high mortality rate of lung disease in the current world,with the development of the society,the decrease of the air quality and the change of living environment,the scale of tuberculosis disease illness is gradually increasing,which becomes one of the major threat to human life and health.Due to it,the prevention,monitoring,diagnosis and treatment of tuberculosis disease becomes a worldwide topic.With the development of the disease,the damage to the lungs is becoming more and more serious,which will cause the erosion and damage of the lungs.CT image has become the key basis for clinical diagnosis and treatment.Segmentation of lung from CT images can detect the degree of lung injury,which is of great significance to assist doctors in more accurate and rapid disease detection.Traditional medical segmentation and detection methods need to be modified artificially and they are inefficient.With the increasing urgency of medical resources,computer-aided doctors to complete automatic segmentation has gradually attracted the attention of the industry and many researchers.Tuberculosis has serious lung erosion,and CT images of tuberculosis lung contain a large number of broken and separated lung parenchyma,which brings great challenges to the automatic segmentation algorithm.However,the existing deep learning methods cannot effectively detect the damaged lung regions and accurately segment the discrete small areas of lung parenchyma when processing the pulmonary CT images of tuberculosis damaged and separated lung.Therefore,in this dissertation,we propose an automatic tuberculosis lung segmentation algorithm based on CT images.The main research work of this dissertation is as follows:Firstly,the CT images are analyzed and preprocessed to build a data set.According to the existing CT sequence file analysis and extraction and the clinical needs of the window adjustment and section selection,image data set is generated;Adaptive denoising and normalization processing are carried out on the generated tuberculosis CT images to improve the contrast between the lung and the background,retain the lesion details,and construct complete and reasonable tuberculosis CT image data for model training and verification.Secondly,an automatic segmentation algorithm for damaged tuberculous lung under multi-scale deep supervision is proposed.In order to adapt to the blur edge of tuberculosis CT images and the characteristics of discrete lung area,we use the multi-scale feature fusion and the dilated convolution induced residual coding module to extract the image features.And in order to correct decoding results,we use a multi-scale deep supervision to make up the loss of pixel in image reconstruction and enhance the segmentation accuracy.Finally,aiming at the problems of the above algorithm,the model is improved,and an automatic segmentation algorithm for damaged tuberculosis lung was proposed based on attention mechanism.To solve the problem of poor segmentation effect of the proposed model on CT images with severe lung invasion and injury,an automatic segmentation algorithm of damaged tuberculosis lung based on dual attention mechanism is proposed.The attention module is designed and introduced to make the results closer to the real gold standard of segmentation.The experimental results show that the model proposed in this dissertation achieves better results than the existing automatic segmentation models in different degrees of disease and different damaged lung images.It can effectively segment the lung details in pulmonary tuberculosis CT with different degrees of damage,which has a better research and application value.