Improved neuroimaging meta analyses with modern text based processing

科学界在很大程度上依赖于元分析，以在迅速增长的文献中建立可信的结果。个别研究受到偏见、报告限制或缺乏力量的影响，所有这些都很难评估。荟萃分析是达成共识并从生命科学和神经科学文献中提取可靠信息的关键。随着科学出版物数量的迅速增加，假阳性结果也在增加。由于神经成像实验的成本很高，研究中参与者的数量往往提供的力量有限。当前的元分析工具，如NeuroSynth或NeuroQuery，能够自动处理从其中提取大脑活动坐标的大型(~14,000)神经成像出版物语料库，并提供一种方法来计算与已标记术语的出版物相关联的脑图。这些工具没有考虑到手稿中使用的方法，尽管已经证明分析管道对神经成像结果有重大影响。文本分析和自然语言处理的先进方法不仅应该允许更好的文献计量学，而且应该改进文献的自动化分析。我们建议探索利用文献的高级文本挖掘分析来改进神经成像中的Meta分析的方法。我们的第一个任务是为大型语料库中的每个出版物自动提取神经成像领域以及数据处理和分析方法的标签或特征向量表示。一旦文章已经用神经成像方法的标签进行了注释，我们建议通过比较使用或不使用该特定方法(例如，簇大小测试、用于功能MRI分析的特定类型的运动校正)的文章的荟萃分析结果来分析给定方法的影响。我们还将研究结果随发表年份的时间演变，评估是否与新方法的出现有联系，并对在荟萃分析中引入偏差和/或方差的神经成像方法提出纠正策略。最后，我们将使用荟萃分析的结果来加强和改进对有限样本量的新研究的分析，并使用一个非常大的神经成像数据集来验证使用荟萃分析图作为先验来正规化统计和预测结果。

The scientific community relies heavily on meta-analyses to establish trustworthy results in a rapidly growing body of literature. Individual studies are subject to biases, reporting limitations, or lack of power, all of which are difficult to assess. Meta-analyses are key to reach consensus and extract solid information from the literature in life sciences and neuroscience. As the number of scientific publications is rapidly increasing, false positive findings are increasing as well. Because neuroimaging experiments are costly, the number of participants included in studies often provides limited power. Current meta-analysis tools -such as Neurosynth or NeuroQuery- are able to automatically process a large (~14,000) corpus of neuroimaging publications from which the brain activity coordinates have been extracted, and provide a way to compute brain maps associated with the publication labelled terms. These tools do not account for the methods used in the manuscript, even though it has been demonstrated that analysis pipelines have a major impact on neuroimaging results. Advanced methods in text analysis and natural language processing should not only permit better bibliometrics but also improved automatized analysis of the literature. We propose to explore ways to improve meta-analyses in neuroimaging with advanced text mining analysis of the literature. Our first task is to automatically extract labels or feature vector representations of both the neuroimaging domains and the data processing and analysis methods for each publication in a large corpus. Once articles have been annotated with labels for neuroimaging methods, we propose to analyse the impact of a given method by comparing meta-analysis results from articles using or not this specific method (e.g., cluster size test, specific type of movement correction for functional MRI analysis). We will also study the temporal evolution of results with the year of publication and assess if there is a link with the emergence of new methods, and propose correction strategies for those neuroimaging methods that introduce bias and/or variance in meta-analyses. Finally, we will use the meta-analysis results to robustify and improve analysis of new studies with limited sample sizes, and employ a very large neuroimaging dataset to validate the use of meta-analytic maps as priors to regularize statistical and prediction results.