基于空洞卷积密集连接网络的左心室MR图像分割方法

Left ventricular MR image segmentation method based on dilated convolution DenseNet

左心室核磁共振(MR)图像分割对于评估心脏功能和诊断疾病具有重要意义.传统分割算法对于左心室,尤其是含有左心室流出道的左心室MR图像,存在分割精度不够的问题.设计了一种基于空洞卷积密集连接网络的左心室MR图像分割方法.该方法利用密集连接网络和空洞卷积缓解了深度学习中梯度消失和内存过度消耗的问题,并且通过数据增强和提取感兴趣区域的方法提升了网络的准确性.分割结果采用平均垂直距离、Dice系数等指标进行评价分析.在MICCAI2009心室分割数据集的138张图片上的测试结果为:内、外膜的平均Dice系数分别为0.91和0.96,平均垂直距离分别为1.71和1.42.实验结果表明,此方法分割精度明显高于其他方法,对于含有左心室流出道的MR图像也能准确分割.

Left ventricular Magnetic Resonance(MR)image segmentation is important for assessing cardiac function and diagnosing disease.The traditional segmentation methods are not accurate enough for left ventricular image,especially for the left ventricular image with outflow tract.In this paper,a segmentation method based on dilated convolution DenseNet has been designed.This method uses DenseNet and dilated convolution to alleviate the problems of gradient disappearance and excessive memory consumption in deep learning,and improve the accuracy of network by data augmentation and extraction of ROI.The segmentation results are evaluated by the Average Perpendicular Distance and the Dice coefficient.The proposed method has been tested on 138 images from the MICCAI2009 ventricular segmentation dataset.The average Dice coefficients of the endocardium and epicardium are 0.91 and 0.96,respectively,and the Average Perpendicular Distance are 1.71 and 1.42,respectively.The results show that the accuracy of the proposed method is significantly higher than other methods,and it can also accurately segment MR image containing the left ventricular outflow tract.