Beijing Institute of Technology, Institute of Signal and Image Processing, School of Information and Electronics, Haidian District, Beijing, 100081, China.
Biomed Eng Online. 2019 Jan 21;18(1):5. doi: 10.1186/s12938-019-0623-8.
Segmenting the hippocampal subfields accurately from brain magnetic resonance (MR) images is a challenging task in medical image analysis. Due to the small structural size and the morphological complexity of the hippocampal subfields, the traditional segmentation methods are hard to obtain the ideal segmentation result.
In this paper, we proposed a hippocampal subfields segmentation method using generative adversarial networks. The proposed method can achieve the pixel-level classification of brain MR images by building an UG-net model and an adversarial model and training the two models against each other alternately. UG-net extracts local information and retains the interrelationship features between pixels. Moreover, the adversarial training implements spatial consistency among the generated class labels and smoothens the edges of class labels on segmented region.
The evaluation has performed on the dataset obtained from center for imaging of neurodegenerative diseases (CIND) for CA1, CA2, DG, CA3, Head, Tail, SUB, ERC and PHG in hippocampal subfields, resulting in the dice similarity coefficient (DSC) of 0.919, 0.648, 0.903, 0.673, 0.929, 0.913, 0.906, 0.884 and 0.889 respectively. For the large subfields, such as Head and CA1 of hippocampus, the DSC was increased by 3.9% and 9.03% than state-of-the-art approaches, while for the smaller subfields, such as ERC and PHG, the segmentation accuracy was significantly increased 20.93% and 16.30% respectively.
The results show the improvement in performance of the proposed method, compared with other methods, which include approaches based on multi-atlas, hierarchical multi-atlas, dictionary learning and sparse representation and CNN. In implementation, the proposed method provides better results in hippocampal subfields segmentation.
从脑磁共振（MR）图像中准确分割海马亚区是医学图像分析中的一项具有挑战性的任务。由于海马亚区的结构尺寸小且形态复杂，传统的分割方法很难获得理想的分割结果。
在本文中，我们提出了一种使用生成对抗网络的海马亚区分割方法。该方法通过构建 UG-net 模型和对抗模型，并通过交替训练这两个模型，实现对脑 MR 图像的像素级分类。UG-net 提取局部信息并保留像素之间的相互关系特征。此外，对抗训练在生成的类别标签之间实现空间一致性，并在分割区域上平滑类别标签的边缘。
在从中心成像神经退行性疾病（CIND）获得的数据集上对 CA1、CA2、DG、CA3、Head、Tail、SUB、ERC 和 PHG 进行了评估，得到的海马亚区的骰子相似系数（DSC）分别为 0.919、0.648、0.903、0.673、0.929、0.913、0.906、0.884 和 0.889。对于较大的亚区，如海马的 Head 和 CA1，DSC 分别比最先进的方法提高了 3.9%和 9.03%，而对于较小的亚区，如 ERC 和 PHG，分割精度分别显著提高了 20.93%和 16.30%。
与包括基于多图谱、分层多图谱、字典学习和稀疏表示以及 CNN 的方法在内的其他方法相比，该方法在性能上有所提高。在实现方面，该方法在海马亚区分割方面提供了更好的结果。
