利用几何信息辅助的超声心动图实时分割

Real-Time Segmentation of Echocardiograms with Geometric Information Assistance

超声心动图的分割在临床上对医生的诊断有巨大作用.针对超声图像含有大量噪声、轮廓特征不明显、已有分割算法耗时久、目标分割不完整或引入不必要的背景区域的问题,提出一种基于全卷积网络与几何信息辅助分割的实时分割算法.首先,利用改进的YOLACT框架并行生成原型模板掩码和左右心室、心房的实例掩码的系数,并将两者线性组合获得实例掩码;然后,利用编码模块增强分割效果,提出位置编码模块避免卷积神经网络带来的全局位置信息丢失,以及提出形状编码模块减少心房心室差异小带来的分类错误.实验结果表明,在超声心动图像数据集上的APA,AMIoU和ADICE指标分别达到0.777,0.705和0.827,该方法比其他算法在精度上接近nnU-Net的结果,但速度可以达到27帧/s,比UNet++提升145%.

Echocardiogram segmentation plays a huge role in the diagnosis of physicians in clinical practice.To address the problems that ultrasound images contain a large amount of noise,contour features are not obvious,and existing segmentation algorithms are time-consuming,incomplete target segmentation or introduce unnecessary background regions,a real-time segmentation algorithm based on full convolutional network with geometric information-assisted segmentation is proposed.First,the improved YOLACT framework is used to generate the coefficients of the prototype template mask and the left and right ventricular and atrial instance masks in parallel,and they are combined linearly to obtain the instance mask.Then the encoder module is used to enhance the segmentation effect,the position encoding module is proposed to avoid the loss of global position information caused by the convolutional neural network,and the shape encoding module is proposed to reduce the classification error caused by the small difference between atria and ventricles.Then,the encoder module is used to enhance the segmentation effect.The experimental results show that that the APA,AMIoU and ADICE metrics on the echocardiographic image reach 0.777,0.705 and 0.827,respectively,and the method obtains results close to nnU-Net in terms of accuracy than other algorithms,but the speed can reach 27 frames per second,which is 145%better than UNet++.