| 基于2.5D级联卷积神经网络的CT图像胰腺分割方法 |
| |
| 引用本文： | 周芳芳,王一达,宋阳,杨光.基于2.5D级联卷积神经网络的CT图像胰腺分割方法[J].中国医学物理学杂志,2020,37(6):786-791. |
| |
| 作者姓名： | 周芳芳 王一达 宋阳 杨光 |
| |
| 作者单位： | 华东师范大学物理与电子科学学院/上海市磁共振重点实验室， 上海 200062 |
| |
| 基金项目： | 国家自然科学基金(61731009,81771816)。 |
| |
| 摘 要： | 目的:由于胰腺体积小、形态个体差异性大,影像上的准确分割较为困难。本文提出一种基于2.5D级联卷积神经网络的CT图像胰腺分割方法。方法:实验中使用的数据为NIH胰腺分割公开数据集,共包含82例腹部CT图像,随机选取其中56、9、17例分别作为训练集、验证集和测试集;训练过程中使用旋转、拉伸、平移、裁剪等操作对数据进行扩增。实验中提出一种用于胰腺分割的、结合概率图的2.5D级联深度监督UNet,即CSNet(Cascading deep Supervision UNet)。该网络由3个部分组成:第1部分基于UNet,输入连续5层图像,输出中间3层对应的粗分割图像,设置适当的阈值,使其变成二值的粗分割结果;第2部分将第1层、第3层的粗分割结果与中间层的原始图像相结合,输入另一个深度监督UNet网络,得到中间层的精细分割;第3部分将第1部分网络输出的中间层的粗分割概率图与第2部分网络输出的细分割概率图通过1×1卷积进行概率融合得到最终的输出结果。3个子网络同时进行训练,对应的能量函数联合优化,从而得到更精准的分割结果。最后,使用DSC对分割结果进行评估。结果:在独立测试集上,CSNet实现了(83.74±5.27)%的DSC值。结论:CSNet可以准确分割出CT图像上的胰腺区域。
|
| 关 键 词： | 卷积神经网络 胰腺分割 深度监督 |
Segmentation of pancreas in CT images based on 2.5D cascaded convolutional neural network |
| |
| ZHOU Fangfang,WANG Yida,SONG Yang,YANG Guang.Segmentation of pancreas in CT images based on 2.5D cascaded convolutional neural network[J].Chinese Journal of Medical Physics,2020,37(6):786-791. |
| |
| Authors: | ZHOU Fangfang WANG Yida SONG Yang YANG Guang |
| |
| Affiliation: | Shanghai Key Laboratory of MRI /School of Physics and Electronic Science, East China Normal University, Shanghai 200062, China |
| |
| Abstract: | Abstract: Objective Due to the small size of the pancreas and inter-individual structural variances, it is difficult to segment the pancreas in images accurately. Therefore, a 2.5D cascaded convolutional neural network is proposed for the segmentation of pancreas in CT images. Methods The data used in the experiment were the open dataset for pancreas segmentation published by NIH. The 82 abdominal CT images contained in the dataset were randomly divided into training, validation, and test sets (training/validation/test=56/9/17). During the training, rotating, stretching, shifting and shearing were used for data augmentation. The combination of probability graph and 2.5D cascaded deep supervision UNet, namely CSNet (Cascading deep Supervision UNet), was proposed for pancreas segmentation. The proposed network consisted of 3 sub-networks. Based on UNet, 5 consecutive slices of image were input into the first sub-network to obtain the coarse segmentation images of the middle 3 slices. After a suitable threshold was set, the segmentation images were transformed into binary segmentation images. The original image at the center slice was combined with the coarse segmentation images of two adjacent slices (the first and third slices) and then input into the second sub-network, another deep supervision UNet, to obtain a fine segmentation of the center slice. In the third sub-network, the coarse segmentation probability images of the middle slices from the first sub-network were combined with the fine segmentation probability images from the second sub-network by 1×1 convolution to obtain the final segmentation. The 3 sub-networks were trained simultaneously. During the training, the energy functions of 3 sub-networks were optimized jointly, thereby obtaining a more accurate segmentation result. Finally, the results of pancreas segmentation were evaluated by DSC. Results CSNet achieve a DSC of (83.74±5.27)% on the independent test set. Conclusion CSNet can be used to accurately segment the pancreas in CT images. |
| |
| Keywords: | Keywords: convolutional neural network pancreas segmentation deep supervision |
| 本文献已被 维普 等数据库收录！ |
| 点击此处可从《中国医学物理学杂志》浏览原始摘要信息 |
|
点击此处可从《中国医学物理学杂志》下载全文 |
|
