Recognition And Evaluation Of CTPA Pulmonary Embolism Based On Deep Learning

Pulmonary embolism is a disease with high morbidity and mortality.It is caused by emboli,which are produced inside the body or originate from outside,and block the main or branch of the pulmonary artery,resulting in poor pulmonary circulation and breathing.Because its clinical features are very similar to other respiratory diseases,it is difficult to diagnose the disease based on clinical manifestations alone.CTPA is the imaging diagnostic method of choice for the detection of pulmonary embolism.Since the contrast agent dissolves in the blood vessels and the pulmonary emboli do not absorb the contrast agent,pulmonary emboli appear as relatively dark areas in bright pulmonary arterial areas on CTPA images.In clinic,the diagnosis of pulmonary embolism is usually made by eye scanning.However,it is time-consuming and laborious to detect pulmonary embolism by visual inspection by a radiologist.The use of computer-aided detection methods to identify pulmonary embolism and evaluate the degree of pulmonary embolism can save time,assist doctors in diagnosing the disease,and improve the accuracy of diagnosis,which is of great research significance.The method based on deep learning in this thesis firstly identifies pulmonary embolism,and then performs the segmentation of pulmonary embolism and pulmonary artery on CTPA sequence images,and calculates the risk assessment coefficient.The main contributions include:(1)The MA Faster R-CNN architecture for pulmonary embolism recognition is proposed.As the pulmonary embolism of the research object was a small object with obscure characteristics,and there were problems of missed detection and misdetection when the original Faster R-CNN was used.To solve the above problems,this thesis proposes a new feature fusion network named Multi-scale Fusion Feature Pyramid Network(MF-FPN).It adds two bottom-up information flows on the basis of the original feature pyramid network structure,which strengthens the entire model’s ability to extract the detailed information and position information of the object.(2)The residual structure is added to the prediction module of MA Faster R-CNN,and a residual prediction module(RPM)is proposed,which increases the depth of the classification network and improves the average accuracy of pulmonary embolism recognition.Compared with the original prediction module,the network with the addition of the residual prediction module has a higher confidence score in the detection of pulmonary emboli.(3)An improved U-Net architecture is proposed.In order to solve the problems of low segmentation accuracy and missing segmentation of original U-Net in pulmonary embolus segmentation,a multi-scale feature fusion module was introduced into the contraction path of U-Net structure.It fuses the feature information of adjacent high-level and low-level layers,and combines the fused features with the features generated by the expansion path,realizes feature reuse,improves the network model’s ability to extract feature,and effectively improves the segmentation effect.Based on the segmented pulmonary emboli and pulmonary arteries,the proportion of the pulmonary emboli area in the pulmonary artery area was obtained,and the degree of pulmonary embolism blockage was evaluated according to the proportion value for the reference of doctors.