Semantic Segmentation In Coronary Angiography And Application In 3D Reconstruction

Cardiovascular disease has attracted widespread attention as the the disease with the highest proportion of deaths among residents.The factors such as unbalanced distribution of social professional medical resources and misdiagnosis from doctors who are under huge work pressure are the underlying social reasons leading to the high proportion of deaths from this disease.With the development of modern storage technology,massive amounts of clinical data have been accumulated in hospitals.It brings opportunities for data-driven deep learning technology.In recent years,deep learning algorithms have gradually become active in various medical image analysis fields to assist doctors in diagnosis.In terms of coronary-related clinical data analysis,vascular segmentation technology and vascular reconstruction technology are currently effective and important technologies to assist doctors in diagnosis.In this thesis,a lightweight transformer-convolution hybrid UNet,SOFTAlignUNet(SOFT-AU),is proposed to focus on coronary artery main branch segmentation in X-ray coronary angiography images.It introduced a feature alignment algorithm to alleviate the feature deviation problem existing in the UNet-like network.The proposed network not only acheieved the best dice score of 91.23%among the cnntransformer structure on our own clinical dataset,but also obtained highly competitive results with lightweight computing requirements on the three public datasets of Synapse,DRIVE and ACDC.In addition,this thesis summarized the methods that improve UNet-like network on medical image segmentation and proved that these ways are not fully applicable to the transformer structures.In the root node detection task,this thesis proposed two different solutions.One is the key point detection method which adopted object detection network to the key point detection task and the other is the semantic segmentation method which converted the root node to the center point of the juction of catheter and main branch.Finally,according to the results of coronary main branch segmentation and root node detection,this thesis proposed a two-stage calibration algorithm for the 3D reconstruction of the coronary main branch centerline,which greatly reduced the reconstruction error and finally achieved an error of 0.519±0.903mm.