School of Technology, PUCRS, Porto Alegre, Brazil.
School of Technology, PUCRS, Porto Alegre, Brazil; School of Medicine, PUCRS, Porto Alegre, Brazil; Brain Institute of Rio Grande do Sul, PUCRS, Porto Alegre, Brazil.
Comput Biol Med. 2018 Apr 1;95:90-98. doi: 10.1016/j.compbiomed.2018.02.005. Epub 2018 Feb 9.
Current state-of-the-art methods for whole and subfield hippocampus segmentation use pre-segmented templates, also known as atlases, in the pre-processing stages. Typically, the input image is registered to the template, which provides prior information for the segmentation process. Using a single standard atlas increases the difficulty in dealing with individuals who have a brain anatomy that is morphologically different from the atlas, especially in older brains. To increase the segmentation precision in these cases, without any manual intervention, multiple atlases can be used. However, registration to many templates leads to a high computational cost. Researchers have proposed to use an atlas pre-selection technique based on meta-information followed by the selection of an atlas based on image similarity. Unfortunately, this method also presents a high computational cost due to the image-similarity process. Thus, it is desirable to pre-select a smaller number of atlases as long as this does not impact on the segmentation quality. To pick out an atlas that provides the best registration, we evaluate the use of three meta-information parameters (medical condition, age range, and gender) to choose the atlas. In this work, 24 atlases were defined and each is based on the combination of the three meta-information parameters. These atlases were used to segment 352 vol from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Hippocampus segmentation with each of these atlases was evaluated and compared to reference segmentations of the hippocampus, which are available from ADNI. The use of atlas selection by meta-information led to a significant gain in the Dice similarity coefficient, which reached 0.68 ± 0.11, compared to 0.62 ± 0.12 when using only the standard MNI152 atlas. Statistical analysis showed that the three meta-information parameters provided a significant improvement in the segmentation accuracy.
目前，全脑和子区域海马体分割的最新方法在预处理阶段使用预分割模板，也称为图谱。通常，输入图像会被注册到模板上，为分割过程提供先验信息。使用单个标准图谱增加了处理脑解剖结构与图谱形态不同的个体的难度，尤其是在老年大脑中。为了在这些情况下提高分割精度，而无需任何人工干预，可以使用多个图谱。然而，注册到许多图谱会导致计算成本增加。研究人员提出了一种基于元信息的图谱预选技术，然后基于图像相似度选择图谱。不幸的是，由于图像相似度过程，这种方法也存在计算成本高的问题。因此，只要不影响分割质量，就希望预选较少的图谱。为了选择提供最佳注册的图谱，我们评估了使用三个元信息参数（医疗状况、年龄范围和性别）选择图谱的效果。在这项工作中，定义了 24 个图谱，每个图谱都是基于这三个元信息参数的组合。这些图谱用于从阿尔茨海默病神经影像学倡议（ADNI）数据库中分割 352 个体积。使用这些图谱中的每一个来分割海马体，并与 ADNI 提供的海马体参考分割进行评估和比较。使用元信息进行图谱选择导致 Dice 相似系数显著提高，从使用标准 MNI152 图谱时的 0.62±0.12 提高到 0.68±0.11。统计分析表明，这三个元信息参数显著提高了分割准确性。
