PDF(Pubmed)
Abstract：
Stroke is a leading cause of disability, and Magnetic Resonance Imaging (MRI) is routinely acquired for acute stroke management. Publicly sharing these datasets can aid in the development of machine learning algorithms, particularly for lesion identification, brain health quantification, and prognosis. These algorithms thrive on large amounts of information, but require diverse datasets to avoid overfitting to specific populations or acquisitions. While there are many large public MRI datasets, few of these include acute stroke. We describe clinical MRI using diffusion-weighted, fluid-attenuated and T1-weighted modalities for 1715 individuals admitted in the upstate of South Carolina, of whom 1461 have acute ischemic stroke. Demographic and impairment data are provided for 1106 of the stroke survivors from this cohort. Our validation demonstrates that machine learning can leverage the imaging data to predict stroke severity as measured by the NIH Stroke Scale/Score (NIHSS). We share not only the raw data, but also the scripts for replicating our findings. These tools can aid in education, and provide a benchmark for validating improved methods.
摘要：
中风是导致残疾的主要原因,磁共振成像（MRI）通常用于急性卒中管理。公开分享这些数据集可以帮助机器学习算法的发展，特别是病变识别，大脑健康量化，和预后。这些算法在大量信息中茁壮成长，但需要不同的数据集，以避免过度拟合到特定的人群或收购。虽然有许多大型公共MRI数据集，其中很少包括急性中风。我们使用扩散加权来描述临床MRI，在南卡罗来纳州北部接纳的1715名个体的流体衰减和T1加权模式，其中1461人患有急性缺血性中风。提供了来自该队列的1106名中风幸存者的人口统计学和损害数据。我们的验证表明，机器学习可以利用成像数据来预测由NIH卒中量表/评分（NIHSS）测量的卒中严重程度。我们不仅分享原始数据，还有复制我们发现的脚本。这些工具可以帮助教育，并为验证改进方法提供基准。
