Efficient Kidney Segmentation in Micro-CT Based on Multi-Atlas Registration and Random Forests
Efficient Kidney Segmentation in Micro-CT Based on Multi-Atlas Registration and Random Forests
复制标题
基于多图谱配准和随机森林的显微 CT 高效肾脏分割
DOI：
10.1109/access.2018.2861418
复制
发表时间：
2018
期刊：
影响因子：
3.9
通讯作者：
He Xiaowei
中科院分区：
文献类型：
--
作者：
Zhao Fengjun;Gao Pei;Hu Haowen;He Xuelei;Hou Yuqing;He Xiaowei
微型计算机断层扫描(Micro-CT)为小鼠肾脏提供了体内高分辨率的临床前成像。然而，由于X射线的剂量相对较低，在Micro-CT成像中准确有效地分割小鼠肾脏仍然具有挑战性。提出了一种基于多图谱配准(MAR)和随机森林(RFS)的Micro-CT图像肾脏分割方法。首先，我们构建了MAR的肾脏概率图，得到了肾脏的初步形状估计。我们获得了基于高对比度器官的MAR变换，然后将肾脏映射到新的Micro-CT图像。其次，我们从概率较低的体素中提取多个特征(包括强度、纹理和上下文特征)，并将这些特征提供给RF分类器。射频的作用是微调MAR后的肾脏边界。最后，结合高概率的初始形状和RF的微调，得到最终的肾脏分割。这些实验是在使用和不使用造影剂(Dataset1和Dataset2)的情况下对小鼠进行Micro-CT成像所获得的数据集上进行的。结果表明，MAR-RF优于水平集、统计图谱配准、主动形状模型和其他有监督学习方法，DICE系数分别为0.9766和0.9255，数据集1和数据集2上的平均表面距离分别为1.2 5和0.98 mm。我们的MAR-RF的训练时间和预测时间分别只有比较方法的37.04%和17.68%。该方法在其他计算机辅助诊断的分割任务中具有很大的应用潜力。
Micro-computed tomography (micro-CT) provides an in vivo high-resolution preclinical imaging for murine kidneys. However, due to the relatively low dosage of X-rays, accurate and efficient segmentation of murine kidneys in micro-CT imaging remains challenging. In this paper, we proposed an efficient kidney segmentation method in micro-CT images based on multi-atlas registration (MAR) and random forests (RFs). First, we constructed a probability map of kidneys by the MAR and obtained an initial shape estimation of kidneys. We acquired the transformations of MAR based on high-contrast organs and then mapped kidneys to the new micro-CT images. Second, we extracted multiple features (including intensity, texture, and context features) from the voxels with lower probabilities and fed these features to a RF classifier. The role of RF is to fine-tune kidney boundaries after MAR. Finally, combining the initial shape with high probabilities and the fine-tuning of RF, we obtained the final segmentation of kidneys. The experiments were conducted on datasets acquired by micro-CT imaging of mice with and without the administration of contrast agent (Dataset1 and Dataset2). The results demonstrated the proposed MAR-RF outperformed the level sets, statistical-atlas registration, active shape model, and other supervised learning methods, with the Dice coefficients of 0.9766 and 0.9255, and the mean surface distances of 1.25 and 0.98 mm on Dataset1 and Dataset2, respectively. The training and prediction time of our MAR-RF were only 37.04% and 17.68% of the compared method, respectively. The proposed method has great potential for applications in other segmentation tasks of computer-aided diagnosis.