研究大脑结构和功能之间的映射关系是认知神经科学的基本问题。随着科技的发展，近年来研究者能够利用功能核磁共振技术对大脑工作机制进行研究。通过实验设计和数据采集分析，从而定位出人类大脑上具有特定功能选择性的大脑区域，例如与面孔识别有关的外侧梭状回中部的梭形面部区域(Fusiform Face Area, FFA)。 为了进一步更深层次的研究这些分布在人类大脑皮层不同部位具有特定功能选择性的脑区，研究者通常在这些已知区域内设计更为复杂的认知实验任务，例如在面孔识别相关区域内进行面孔记忆机制的研究。这类方法被称为功能感兴趣区域分析。在感兴趣区域分析中，首先需要在被试的功能激活图中确定感兴趣区域的范围。在以往大量的研究中，确定大脑功能感兴趣区域的方法主要是通过专家手动标注的方式完成的。然而，手动标注的方法过于耗时，在数据量较大的情况下成本过高。同时，大脑中的功能区域在被试之间具有较大个体差异，人工手动标注的方法容易遭受主观因素的影响，这些因素可能会给实验结果带来一定偏差。本研究将采用机器学习的算法框架解决功能感兴趣区域标注的问题。 研究中将功能感兴趣区域的标注任务转化为大脑激活体素的多分类问题。采用随机森林的分类算法，抽取多通道核磁影像中体素级别的图像特征，训练分类器并进行预测，对大脑功能感兴趣区域进行标注。研究中基于面孔识别相关功能区域进行了实验，结果显示在四个面孔识别相关的功能区域上平均Dice相似度指标达到84%。相比以往的组约束的方法，本研究采用的方法能够准确，客观和自动地完成大脑功能感兴趣区域的标定，方法充分的尊重了大脑功能区的个体差异。进一步，实验中验证了图像上下文特征，多通道的核磁图像信息和大的样本量都能有效地提高标注的准确度。

Mapping function in human brain is a main goal of cognitive neuroscience. With the devel-opment of science and technology, experts could use fMRI experiments to explore the brain functional structure or mechanism. As we know, a lot of functional selective regions have been found, like Fusiform Face Area (FFA). Actually, experts could define functional regions of interest (fROI) for further determine more elaborate cognitive function of those distributed functional brain regions. This work is usu-ally by hand picking. However, it would cost a lot of time, when we have too many subjects. At the same time, the subjective nature of the defining process and the variability across individ-uals in fROI locations may cause bias when make the decision. In this paper, we employ machine learning framework to handle the labeling task. In this study, we turn the labeling of fROI into multi-classification problem. Random forests has been employed to define the fROI automatically with voxel-wise and context-awareness feature types from multiple MRI image channels. We have conducted a series of experiments on four face-selective functional regions, the result shows that our method could obtains an aver-aged Dice similarity score of 84%. We compare the performance of our method with GSS(Group-Constrained Subject-Specific) method, the result shows that our method could han-dle the functional labeling of fROI more accurately, objectively and automatically and shows more respect to the individual difference of fROI. We also prove the context-awareness feature type, structure brain signal intensity and the large training dataset used is helpful to improve the classifier performance.