结合Transformer和CNN的超声心动图左心室分割算法

Left Ventricular Segmentation Algorithm for Echocardiography Combining Transformer and CNN

针对超声心动图因噪声大、左心室边缘模糊造成分割困难和分割结果不准确的问题,提出一种结合Transformer和CNN的左心室分割算法.首先利用2种网络构架高效地捕捉全局特征和局部细节;其次使用卷积注意力设计特征融合模块,自适应地融合来自Transformer和CNN分支的特征;最后引入桥注意力模块并根据3层融合特征计算注意力特征图,得到更精确的分割结果.在大型心脏图像数据集EchoNet-Dynamic上进行训练、验证和测试的结果表明,所提算法的Dice系数达到92.41%,分割性能优于6种对比算法;在临床患者的超声图像上,可视化和临床医生的盲审结果证明了该算法的有效性.所提算法可以精确地分割左心室,为心脏疾病诊断提供可靠的计算机辅助.

To address the problem of difficult and inaccurate segmentation of left ventricle due to high noise and fuzzy edge in echocardiography,we propose a left ventricle segmentation algorithm that combines Transformer and CNN.First,we use two network architectures to efficiently capture global features and lo-cal details.Second,we design a feature fusion module using convolutional attention to adaptively fuse fea-tures from Transformer and CNN branches.Finally,we introduce a bridge attention module and calculate attention feature maps based on three-layer fusion features to obtain more accurate segmentation results.To validate the performance of HeartNet,we train,validate and test it on a large-scale cardiac image dataset EchoNet-Dynamic,achieving a Dice Coefficient of 92.41%,which outperforms six other algorithms in-volved in comparison.We test it on clinical patients’ultrasound images,and the visualization and blind re-view results by clinical doctors demonstrate the effectiveness of this algorithm.The experimental results show that HeartNet can accurately segment left ventricle,providing reliable computer assistance for cardiac disease diagnosis.