Research On Chronic Obstructive Pulmonary Disease Classification In CT Image Based On Deep Learning

Chronic obstructive pulmonary disease(COPD)has a high morbidity and mortality rate worldwide,early detection and treatment of COPD is an important approach to improve patient survival rate and prognosis.Chest computed tomography(CT)detection has emerged as an effective method that can be used to clinically quantify and diagnose COPD.Whereas many CT-based deep learning approaches have been developed to identify COPD,it remains challenging to characterize the highly spatially heterogeneous disease features of COPD based on CT and the diagnosis performance still needs to be improved due to the high complexity of COPD pathology and the lack of accurate annotation of lesions in its CT images.Therefore,based on COPD CT images,this project conducted research on COPD image classification problems based on deep learning.The main research contents of the project are as follows:(1)A feature fusion and improved 3D-DenseNet-based method is proposed to identify COPD using CT images.This method firstly preprocesses the CT images,and then extracts the bilateral lung region and airway tree region in the CT images using a fully automated lung tissue segmentation algorithm;subsequently,considering the 3D characteristics of CT images and in order to improve the recognition accuracy and generalization ability of the model,an improved 3D-DenseNet network architecture is constructed by replacing the Dropout layer with the DropBlock module,and the final classification model is obtained by using the fusion images of bilateral lung CT and airway tree CT as input in training.The experimental results show that the problem of lack of accurate annotation of COPD CT images can be effectively solved by training the model with 3D convolutional neural network,and the improved 3D-DenseNet architecture also has better regularization effect,which can improve the accuracy and generalization of COPD diagnosis model.(2)A multiple instance learning(MIL)with two-stage attention(TSA-MIL)is proposed to identify COPD using CT images.This method firstly constructs data bags and instances required for MIL through the steps of CT image preprocessing,lung region segmentation and CT slice selection;secondly,to alleviate the lack of training data,the pre-trained ResNet-50 is used to extract multicomponent and multidimensional features of COPD abnormalities through parameter-transfer learning,in which a pseudo-color image processing method is designed to transfer single-channel CT slices to RGB-like three channels by using window processing technology,which makes the converted CT images satisfy the input of the transfer learning network on the one hand,and increases the richness of its information carrying on the other hand;subsequently,a two-stage attention mechanism module is constructed,in which the first stage aiming at discovering the key instance while the second stage correcting the attention score for each instance by calculating its average relative distance to the key instances,which can avoid the computational error caused by a single attention operator and thus generate more robust attention scores for each instance;finally,the attention scores are weighted and aggregated for all instances to obtain bag-level embedded feature representations and obtain the final prediction results;besides,an instance-level clustering over feature domain is exploited to further improve feature separability and therefore improve the classification accuracy of COPD.The results of a large number of comparison experiments and ablation experiments show that the CT image classification model based on MIL with two-stage attention can not only be trained quickly,but also can accurately identify patients with COPD.More importantly,the model can be applied to small sample datasets,and to some extent,it can alleviate the common multi-center effect problem in the field of medical image classification.