Research On Detection Method Of Stenosis Area In Coronary Angiography Images

Coronary artery disease(CAD)is the narrowing of the lumen of vessels mainly caused by atherosclerosis or spasm of the coronary arteries,which is one of the leading causes of death.X-ray coronary angiography(XCA)is the gold standard imaging technique for the diagnosis of CAD.It is used both as a reliable basis for preoperative assessment of coronary stenosis and for monitoring the progress of revascularization during angioplasty.In clinical practice,determination of coronary stenosis and its severity is done manually by physician visual assessment(PVA).Due to differences in patient condition and image quality,as well as complex coronary structures,there is inter-rating variability in diagnostic results.And a lot of observation will cause the doctor’s visual fatigue,causing misdiagnosis and missed diagnosis.The existing workflow methods based on computer automatic detection technology have problems such as cascaded error accumulation,poor robustness and time-consuming.Therefore,this study applied deep learning technology to the automatic analysis of coronary angiography images to assist doctors to detect stenosis reliably,including the selection of key frames based on coronary artery segmentation and the identification and localization of coronary artery stenosis.The main research contents of this paper are as follows:(1)Due to the complex diversity of medical image processing and analysis tasks,non-standardized acquisition schemes and isolation of medical images,centralized public coronary angiography image dataset is rare.In this regard,this study built a segmentation dataset and a stenosis dataset on coronary angiography images,and designed a two-stage detection model of semantic segmentation and narrow detection,focusing on the following problems in stenosis detection: strong subjectivity in key frame selection,complex background noise,soft tissue overlap,low contrast,uneven illumination,and low detection rate of small coronary branch stenosis.The experimental results show the effectiveness of the designed method.At the same time,the model inference speed can meet the requirements of real-time detection,which can ensure that the method can be used to monitor the progress of vascular reconstruction in real time during angioplasty.(2)The selection of key frames in XCA sequences is the premise for further identification and interpretation of coronary angiography images.These key frames filled with contrast agents can completely display the structure and morphology of coronary arteries,which is helpful for subsequent extraction and detection of stenosis features.Aiming at the problem that the current key frame selection mainly relies on manual labor,which is affected by subjective factors and has a large workload,we devised a method to select key frames to control the quality of keyframe selection effectively: first segment the coronary tree using a semantic segmentation model,and then compare the filling degree of the segmented vessel regions.And these key frames are used as input for stenosis detection.At the same time,the segmented blood vessels can also lay the foundation for medical angiodynamic studies and 3D reconstruction.The segmentation effect and the accuracy and inference speed of the experimental results of key frame selection show the effectiveness of the method.The comparison experiment with traditional image processing methods shows that the method is more robust.(3)A stenosis detection network SD-RFH-YOLO(Stenosis Diagnosis and Rich Feature Hierarchies Based YOLO)was designed to identify and locate coronary stenosis to assist doctors in objective interpretation of coronary angiography images.First,by combining channel shuffling and grouping convolution operations,the channel grouping association mechanism module SFPP(Spatial Shuffle Pyramid Pooling)is designed in the backbone layer of the network to solve the problem of information exchange between the related channels of different scale feature maps at the same level in the traditional spatial pyramid pooling layer.And the selflearning space fusion mechanism module Bi-PAN(Bi-directional Path Aggregation Network)is designed in the neck layer to strengthen the fusion of low-level detail features and high-level semantic features.At the same time,the attention mechanism is introduced to assign weights to the input features to represent the contributions of different levels of features,so as to solve the problem of indiscriminate treatment of input features when fusing features.The experimental results show that the network can improve the detection ability of small and medium targets while reducing the amount of calculation.In addition,for false positive results due to contrast agent flow and vascular motion,etc.,a Seq-FPS(Sequence False Positive Suppression)module was designed to suppress false positive results by exploiting the latent temporal properties of XCA sequences.