Dataset of magnetic resonance images of nonepileptic subjects and temporal lobe epilepsy patients for validation of hippocampal segmentation techniques.
Dataset of magnetic resonance images of nonepileptic subjects and temporal lobe epilepsy patients for validation of hippocampal segmentation techniques.
复制标题
DOI：
10.1007/s12021-010-9096-4
复制
发表时间：
2011-12
期刊：
影响因子：
3
通讯作者：
Soltanian-Zadeh H
中科院分区：
文献类型：
--
作者：
Jafari-Khouzani K;Elisevich KV;Patel S;Soltanian-Zadeh H
海马体已成为几种神经退行性疾病研究的焦点。从大脑的磁共振成像扫描中自动分割这种结构有助于这项工作。分割技术必须使用人工生成的精确海马轮廓的MR图像数据集进行评估。手动分割不是一项简单的任务。缺乏独特的分割方案和较差的图像质量是混淆比较研究所需的一致性的两个因素。我们开发了一个公开可用的癫痫和非癫痫受试者的t1加权（T1W） MR图像数据集，以及他们的海马轮廓，以提供一种评估分割技术的手段。该数据集包含50张T1W MR图像，40张癫痫和10张非癫痫。所有图像都通过广泛使用的协议进行手动分割。选择25幅图像进行训练，并给予海马标记。另外25张没有标签的图片用于测试算法。允许用户使用11个分割相似度度量来评估他们为测试图像生成的标签。使用该数据集，我们评估了两种分割算法，脑解析器和分类器融合和标记（CFL），由训练集训练。对于Brain Parser，测试集的平均Dice系数为0.64。对于CFL，该值为0.75。这些发现表明，需要进一步改进分割算法，以提高可靠性。
The hippocampus has become the focus of research in several neurodegenerative disorders. Automatic segmentation of this structure from magnetic resonance (MR) imaging scans of the brain facilitates this work. Segmentation techniques must be evaluated using a dataset of MR images with accurate hippocampal outlines generated manually. Manual segmentation is not a trivial task. Lack of a unique segmentation protocol and poor image quality are only two factors that have confounded the consistency required for comparative study. We have developed a publicly available dataset of T1-weighted (T1W) MR images of epileptic and nonepileptic subjects along with their hippocampal outlines to provide a means of evaluation of segmentation techniques. This dataset contains 50 T1W MR images, 40 epileptic and 10 nonepileptic. All images were manually segmented by a widely used protocol. Twenty five images were selected for training and were provided with hippocampal labels. Twenty five other images were provided without labels for testing algorithms. The users are allowed to evaluate their generated labels for the test images using 11 segmentation similarity metrics. Using this dataset, we evaluated two segmentation algorithms, Brain Parser and Classifier Fusion and Labeling (CFL), trained by the training set. For Brain Parser, an average Dice coefficient of 0.64 was obtained with the testing set. For CFL, this value was 0.75. Such findings indicate a need for further improvement of segmentation algorithms in order to enhance reliability.