Deep Learning-based Object Detection Of Calcified Lesions And Its Application To Single-mode Image Alignment

Coronary artery disease can directly affect a patient’s heart and is extremely dangerous.As one of the high-risk factors,coronary artery calcification can induce blood clots and other life-threatening conditions.With the widespread use of computer technology and deep learning,the use of computer-assisted medicine has become one of the main tools of clinical testing in modern medicine.Coronary angiography has been widely used as the "gold standard" for the diagnosis of coronary artery disease,but there are few studies of calcified lesions based on coronary angiography data.The difficulty of detecting calcified lesions is greatly increased by the influence of vascular branches,background structures and other pathological phenomena in coronary angiography data.Therefore,studies related to calcified lesions using imaging data are of high practical importance.The purpose of this thesis is to design an improved calcified lesion detection algorithm with high detection accuracy based on coronary angiography data,and to further improve its usefulness in aiding diagnosis and treatment using techniques such as image alignment.First,this thesis proposes an MCascade R-CNN network based on an extensive investigation of existing detection networks and the design of balanced aggregation pyramids and convolutional attention detectors.The network is enhanced to detect calcified lesions by enhancing the features of different layers in the feature map and improving the performance of the detector.Experimental results demonstrate that our network has better detection accuracy than existing object detection networks in the detection of calcified lesions,and achieves a detection accuracy of 90.6%and 88.2%recall in the detection of heavy calcified lesions.Second,this thesis designs a single-mode image alignment technique using image segmentation technique to extract interventional catheters for image alignment.After automatically extracting the frames to be aligned using ECG signals,we used PSPNet and BASNet dual networks for catheter segmentation and integrated the image center-of-mass,root node and centerline information to calculate the offset for image alignment.The results show that our normalized mutual information of the aligned image is 0.589,which can assist the physician to obtain more information to some extent.Finally,an end-to-end processing scheme is implemented in both parts of this thesis to simplify the related processes and better achieve the research objectives of this thesis.