Abstract:

This paper proposes a method for active contour segmentation of CT images based on super pixel and CNN. Firstly, the CT image is meshed and assigned a label by the super pixel SLIC method. Secondly, the meshed image is used as a dataset training CNN network to segment the super pixels of organs （such as liver, lung, etc.） and the super pixel’s seed point is connected to become the coarse segmentation boundary. Finally, the rough segmentation boundary is used as the initial contour, and the fuzzy active contour segmentation is performed to obtain the boundary of the organ in the CT image. After a large number of experimental comparisons, the segmental mean DSC coefficient of the lung CT image in this method reacheds 97%, and the average ASD coefficient reaches 1.23 mm. Compared with the reference algorithm in the liver CT image, the average VOE coefficient is reduced by 1% on the premise of ensuring the segmentation accuracy, and the segmentation time of the slice image is increased by 10 seconds on average.

Key words: CT image, super pixel, Convolutional Neural Network（CNN）, fuzzy active contour segmentation

摘要：

提出了一种融合超像素和CNN的CT图像器官主动轮廓分割方法。用超像素SLIC方法将CT图像网格化并分配标签；将网格化后图像作为数据集训练CNN网络分割出器官（如肝脏、肺部等）边界超像素，并将这些超像素的种子点连接成为粗分割边界；将粗分割边界作为初始轮廓，进行模糊主动轮廓分割得到CT图像中器官的边界。经过实验对比，该方法对肺部CT图像的分割平均DC系数达到97%、平均ASD系数达到1.23 mm。在肝脏CT图像方面与参考算法进行相比，在保证分割精度的前提下，VOE系数平均减少1%，切片图像的分割时间平均提高10 s。

关键词: CT图像, 超像素, 卷积神经网络（CNN）, 模糊主动轮廓分割