基于Octave卷积的超声心动图左心室分割方法

Segmentation method of left ventricular echocardiography based on Octave convolution

针对传统卷积神经网络(CNN)对左心室的分割精度低且在特征提取过程中存在特征冗余的问题,在传统卷积神经网络的基础上提出基于Octave卷积的超声心动图左心室分割方法。首先,使用Octave卷积对图像进行特征提取,将特征图分为高频部分和低频部分,在卷积的过程中减少了低频信息的使用,从而降低了网络模型的计算量;其次,提出了新的损失函数,将交叉熵和Dice系数进行加权结合。实验结果表明,利用该方法在二腔心数据集上测试,其分割结果的平均像素交并比(MIoU)能够达到79.21%,较传统的U-net卷积神经网络精度提升6.1个百分点,在拥有低计算量的同时提高了分割精度。

Aiming at the problem that traditional Convolutional Neural Network(CNN)has low segmentation precision of left ventricle and feature redundancy in feature extraction process,based on traditional convolutional neural network,left ventricular echocardiography segmentation method based on octave convolution was proposed. Firstly,octave convolution was used to extract features from the images and divide the feature map into high frequency part and low frequency part,which reduces the use of low frequency information in the process of convolution,thus reducing the computational complexity of the network model. Secondly,a new loss function was proposed,which combined the cross entropy and the Dice coefficient. The experimental results show that the Mean Intersection of Union(MIoU)of the segmentation result can reach79. 21% by using this method on the two-cavity dataset,which is 6. 1 percentage points higher than that of the traditional U-net convolutional neural network. The proposed method has high segmentation accuracy and low computational complexity.