基于深度学习的非零水平集保凸的左心室分割

Left ventricle segmentation based on non-zero level set preserving convexity

摘要:

采用距离正则化水平集演化(distance regularized level set evolution,DRLSE)模型对左心室(left ventricle,LV)进行分割会使其产生锯齿状,分割效果较差。为了解决左心室分割目前面临的问题,本文首先使用基于卷积神经网络(convolutional neural network,CNN)的心肌中心线检测算法,取代了水平集方法的人工初始化过程,其次提出了一种基于非零水平集保凸的左心室分割方法。将距离正则化水平集演化(distance regularized level set evolution, DRLSE)模型(水平集方法)、深度学习方法与新方法的平均度中心性进行比较发现,新方法在收缩末期(end-systole,ES)的平均DC(dice coefficient)值为0.93,高于其他方法；除此之外,新方法在舒张末期(end-diastole,ED)与ES阶段的平均豪斯多夫距离(Hausdorff distance,HD)分别为2.51、2.54,明显小于深度学习方法以及水平集方法。实验结果表明,新方法能够有效地提高分割精度。

Abstract:

Segmentation of the left ventricle(LV) using the distance regularized level set evolution(DRLSE) model causes it to be jagged and poorly segmented.To solve these problems currently faced by LV segmentation,this paper firstly uses a convolutional neural network (CNN)-based myocardial center-line detection algorithm to replace the manual initialization process of the level set method,and secondly proposes a non-zero level set-based preserving convexity LV segmentation method.Comparing the mean degree centrality of the DRLSE (level set method),deep learning method and the new method,it is found that the DC (dice coefficient) of the new method at the end-systole (ES) is 0.93,which is higher than the other methods.In addition,the mean Hausdorff distance (HD) of the new method at the end-diastolic (ED) and ES phases are 2.51 and 2.54,respectively,which is significantly smaller than those of the deep learning method and the level set method.The experimental results show that the new method can effectively improve the segmentation accuracy.