Deep Learning Based Detection And Classification Of Bronchiectasis

Bronchiectasis is a common disease with a long and painful course,which poses a serious threat to people’s life.It is a hotspot and trend of current research to use deep learning technology to detect and classify medical images.However,when using lung CT images for deep learning model to detect and classify bronchiectasis,there are some problems such as less data,small lesions and complex image features.To solve these problems,this thesis studies the detection and classification of bronchiectasis based on deep learning.Deep learning technology is used to learn the characteristics of bronchiectasis in CT images,automatically detect the position of bronchiectasis and classify and score the severity of bronchiectasis,so that it can be used as a very useful reference for doctors to distinguish bronchiectasis,which not only greatly improves the efficiency but also reduces the burden of doctors.The main work of this thesis is as follows:1.Collation and annotation of data sets and data enhancement using improved self-training methods and ACER method is proposed for image enhancement.CT images of bronchiectasis patients were obtained from a physical examination screening project for screening,sorting and fine labeling.In the data enhancement stage,the traditional data enhancement method,GAN data enhancement method and self-training semi-supervised data enhancement method are compared.Finally,the self-training method most suitable for this study is adopted and improved to enhance data and expand the data set.In the image enhancement phase,ACER image enhancement method was proposed,which combined Retinex algorithm with adaptive contrast enhancement(ACE)algorithm to effectively improve the contrast of lung CT images,and experiments were designed to verify its improvement effect on deep learning model detection and classification of bronchiectasis.2.The maximum threshold method was used to segment lung parenchyma and the RDU-Net model was proposed to segment lung lobe.First,the maximum threshold method was used to segment lung parenchyma.Then,the disadvantages of the traditional lung segmentation method and the U-NET model were analyzed,and the RDU-NET combined with Residual module and Dilation Convolution was proposed to complete the lung segmentation task better.Then,the effects of deep learning model on the detection and classification of bronchiectasis with and without lung segmentation were compared,and it was verified that the application of RDU-NET for lung segmentation could improve the detection and classification of bronchiectasis with deep learning model.3.HDC Mask R-CNN model was proposed to detect and classify bronchiectasis.Firstly,the detection and classification effects of one-stage and twostage deep learning models on bronchiectasis were compared.Select the Mask R-CNN model with the best Effect and propose an HDC Mask RCNN model based on Hybrid Dilated Convolution(HDC)structure to increase receptive field and avoid Gridding Effect.Then,an experiment was designed to verify its effect on the detection and classification of bronchiectasis.4.Design of bronchiectasis detection and scoring system based on Wechat applet.After completing all the data processing and model training,we developed and designed a bronchiectasis detection and severity scoring system based on Wechat applet.Based on the Wechat applet,the preprocessing results of selected lung CT images and the results of bronchiectasis detection and severity classification are displayed on the Wechat applet to achieve visualization.The system finally obtained in this thesis can not only automatically detect the location of bronchiectasis in CT images,but also score the severity of bronchiectasis according to the different pulmonary lobes where the bronchiectasis is located,and display the results on Wechat applet of mobile phone.The subsequent development of this thesis will gradually push into practical application,effectively help doctors to judge the patient’s bronchiectasis.