Segmentation And Lesion Classification Of Coronary Angiography Based On Deep Learning

Stenosis or obstruction and calcification of the vascular lumen,leading to myocardial ischemia,hypoxia or necrosis will cause coronary artery disease.Coronary angiography is the most common method for disease observation and diagnosis.However,it takes times and efforts to diagnosis,for it requires professional doctors to work together.This thesis intends to use the deep neural network to assist the diagnosis of coronary angiography,setting the task of coronary artery segmentation and identifying the lesions contained in the video.The main contributions of this thesis can be summarized as three points:Firstly,a high-quality the coronary angiography semantic segmentation dataset and a video lesion classification dataset are built for research.The semantic segmentation dataset contains more than 10,000 sets of pictures and annotations,and the video lesion classification dataset contains about 5000 DICOM digital subtraction videos.Secondly,this thesis improves the asymmetric convolution,designs a neural network of the codec structure using a gate structure to adaptively adjust the features that transfer between encoder and decoder;applies structure similarity loss function and a loss function based on Level-Set theory to improve the prediction accuracy;also a post-processing method based on the particle flow theory is proposed and realized to optimize the segmentation continuity.Thirdly,this thesis designs a video classification network based on convolutional LSTM,and proposes a spatial-temporal attention module aiming to make the network focus on the effective parts of time domain and space domain in the coronary angiography video,further optimizing the hidden state extracted;auxiliary loss function is proposed to constrains the distribution of attention weights and speeds up the training convergence.After numerous comparisons and ablation experiments,the effectiveness and practicability of the modules and networks proposed in this thesis have been proved.Some experiments used public datasets for verification and have excellent performance.This thesis completes the task of segmentation of coronary artery and identifying the lesion in real-time and precisely.In the end,F1 score of segmentation achieves 84%,and video lesion mAP achieves 94%,which reaches the advanced level,also the validity of the research content in this thesis is proved qualitatively from clinical medicine,and the research in this thesis is also used in clinical trials.