【摘要】 冠状动脉造影是诊疗冠心病等心血管疾病的一种重要手段，快速而准确的血管分割对诊疗心血管疾病具有十分重要的意义。针对现有冠状动脉造影血管分割算法对细微血管的分割能力不强、分割血管的连通性较差、抗噪声及伪影能力弱等问题，本研究吸取了Transformer结构长距离依赖与跨域跳转连接的优点，分别采用上下文分层聚合和多尺度特征融合的方法，对U型分割网络进行改进，称HAM-UNet。首先，采取必要的图像预处理方法，对原有的冠脉造影图像进行一些特征强化，并扩大了实验数据；然后，将预处理好的图片以HAM-UNet的方法进行分割。编码器同时结合深度卷积与残差结构，可以高效的捕获全局特征并有效增强网络细节感知力，提升分割精度的同时提高分割连通性。解码器进行了多尺度的特征融合，并且加入上采样跳转连接，网络的全局感知得到提高，有效降低了无关信息的影响。所使用数据集来自于天津市医科大学总医院的221张图像和秦皇岛市第一医院的494张图像，在两个数据集上，HAM-UNet算法的准确率分别为0.983和0.998,IOU分别为0.857和0.908,Dice分数分别为0.842和0.883;综合分割性能比U-Net和Att-UNet等算法有较大提升。更多还原

【Abstract】 Coronary angiography is a significant diagnostic and therapeutic modality for coronary heart disease and other cardiovascular diseases. The accurate and expeditious segmentation of blood vessels is of paramount importance to the diagnosis and treatment of cardiovascular diseases. Existing coronary angiography vessel segmentation algorithms have been shown to have several shortcomings, including a weak segmentation ability for fine vessels, poor connectivity of segmented vessels, and a lack of resistance to noise and artefacts. This study proposes an enhanced U-shape segmentation network, termed HAM-UNet.UNet, which utilises the advantages of the Transformer structure′s long-distance dependence and cross-domain hopping connectivity. The proposed methods include contextual hierarchical aggregation and multiscale feature fusion. Firstly, a series of image preprocessing methods are employed to enhance certain features of the original coronary angiography images and expand the experimental data. Then, the preprocessed images are segmented by the HAM-UNet method.The encoder combines both deep convolution and residual structure, which can efficiently capture global features and effectively enhance the network detail perception, thus improving the segmentation accuracy while increasing the segmentation connectivity. The decoder performs multi-scale feature fusion and up-sampling hopping connections, improving the global perception of the network and reducing the influence of irrelevant information.The datasets used are from 221 images from the General Hospital of Tianjin Medical University and 494 images from the First Hospital of Qinhuangdao City. On both datasets, the HAM-UNet algorithm achieves an accuracy of 0.983 and 0.998, respectively. As demonstrated in Figure 8, the IOUs are 0.857 and 0.908, and the Dice scores are 0.842 and 0.883, respectively. This indicates that the overall segmentation performance is superior to that of U-Net, Att-UNet and other algorithms.更多还原