【摘要】 目的 钙化斑块的精准分割在准确评估冠状动脉狭窄程度方面具有重要意义。然而，由于冠脉计算机断层成像中的钙化斑块具有尺度小、边界模糊以及与其他组织高度重叠的衰减范围等特点，其精准分割面临诸多挑战。为了解决从冠脉计算机断层成像中分割钙化斑块存在的假阳率高、精度不足等问题，本研究提出了一种结合边缘信息增强和通道注意力的钙化斑块分割方法。方法 首先，构建一个编解码的网络结构，在编码器部分，采用3×1和1×3的卷积块来增强网络对边缘信息的提取能力。同时，在编码器和解码器之间引入了通道注意力模块，使网络能够自适应地关注感兴趣的特征，抑制不必要的血管和背景信息，从而提高钙化斑块分割的准确性。此外，为了降低心外钙化组织误判导致的假阳性，利用预训练的3D U-Net生成心脏掩码，对钙化斑块分割结果进行后处理。最后，从80例数据中筛选出4 826张图像对网络进行训练，并在10例测试集上进行消融实验和对比实验。结果 在测试集上，本研究所提出的方法表现优异。Dice系数为0.867,Jaccard系数为0.769,精确率为0.846,均优于所对比的先进的医学图像分割方法。结论 本研究所提出的方法能够有效地提高钙化斑块的分割精度，对辅助评估冠脉狭窄程度具有潜在的应用价值。更多还原

【Abstract】 Objective Accurate segmentation of calcified plaque is of great significance in accurately assessing the degree of coronary artery stenosis. However, due to the characteristics of small scale, blurred boundaries, and attenuation ranges that highly overlap with other tissues, the accurate segmentation of calcified plaques in coronary computed tomography faces many challenges. To solve the problems of high false positive rates and insufficient accuracy in calcified plaque segmentation from coronary computed tomography angiography, this study proposes a calcified plaque segmentation method that combines edge information enhancement and channel attention. Methods First, a codec network structure was constructed. In the encoder part, 3×1 and 1×3 convolution blocks were used to improve the network’s ability to extract edge information. At the same time, a channel attention module was introduced between the encoder and decoder, allowing the network to adaptively focus on features of interest and suppress unnecessary blood vessels and background information, thereby improving the accuracy of calcified plaque segmentation. In addition, to reduce the false-positive due to misclassification of extracardiac calcified tissue, a pre-trained 3D U-Net was used to generate cardiac masks and post-process the calcified plaque segmentation results. Finally, 4 826 images were selected from 80 cases of data to train the network, and ablation and comparison experiments were performed on 10 cases of test sets. Results On the test set, the method proposed in this study performed excellently. Specifically, the Dice coefficient was 0.867,the Jaccard coefficient was 0.769,and the accuracy was 0.846,which was better than the compared advanced medical image segmentation methods. Conclusions The method proposed in this study can effectively improve the segmentation accuracy of calcified plaques and has potential application value in assisting in the assessment of the degree of coronary stenosis.更多还原