功能性核磁共振实验事件相关设计的双阶段优化法  

Optimization of Event-Related Design in functional Magnetic Resonance Imaging:A Two-stage Approach

作　　者：莫测[1] 

机构地区：[1]华南师范大学心理学院应用心理研究中心,广州510631

出　　处：《心理科学》2013年第6期1282-1290,共9页Journal of Psychological Science

基　　金：国家重点基础研究发展计划项目(2012CB720700);国家自然科学基金资助项目(30870783)的资助

摘　　要：在使用事件相关设计的功能性核磁共振实验中,为了对实验设计的事件时间序列进行优化以得到最佳的实验结果,目前学术界提出了两种方法:第一种是概率分配法,即通过使不同长度的事件间隔比例服从某个已知的概率分布提高实验效率。这种方法优点是在实现优化的同时保证不同实验条件之间以及不同扫描阶段之间事件间隔长度分布的一致性,但不能稳定地得到最优化的效果。第二种是启发式搜索法,即通过计算机科学中的遗传算法搜索效率最高的事件时间序列。这种方法优点在于可以稳定地获得最优化的效果,但不能保证不同实验条件下以及不同扫描阶段之间事件间隔长度概率分布的一致性,因此在较多的实验研究中是不适合使用的。本文提出基于概率分布和启发式搜索相结合的"双阶段优化法":第一阶段,根据选定的概率分布确定各长度事件间隔的比例,并据此生成大量初始事件时间序列。第二阶段,对前面生成的序列中进行启发式搜索,求出最优解。为验证该方法的可行性,本研究模拟100组fMRI研究的实验设计数据,分别用三种优化方法进行优化并比较优化效果。结果表明,双阶段优化法的实验设计优化效果显著优于概率分配法,且接近于启发式搜索法。据此,本文提出,在事件相关设计中,如果事件间隔长度概率分布的差异对实验结果影响较小,可以采用启发式搜索法;但是,在多数情况下,建议采用双阶段优化法,在确保控制不同实验条件事件间隔长度概率分布这一无关因素的效应的前提下实现设计的最优化。Recent years have witnessed the increasing popularity of event-related design in functional Magnetic Resonance Imaging( fMRI) studies for its desirable ability of constructing counterbalanced trial sequence and remarkable flexibility in statistical modeling.However,event-related designs require more subjects or longer scanning time than blocked designs in order to achieve comparable statistical power due to the more pronounced susceptibility to signal non-linearity and noise temporal autocorrelation. Previous studies have shown that a critical factor constraining the design efficiency of event-related fMRI experiments is the timing of events. Hence,a fundamental challenge in improving design efficiency is the optimization of temporal arrangement of experimental events. It has been proposed that design efficiency can be improved by using designs in which the inter-stimuli-interval( ISI) length conforms to a known probability distribution. In this approach,a probability density function is first selected based on the specific objectives and require-ments of the study to determine the number of occurrence for each ISI type. Although ISI length distributions are well matched across all scanning sessions and all experimental conditions in designs derived via this method,the major drawback lies in its inability to distinguish the best designs among multiple sub-optimal ones. An alternative approach employs the genetic algorithm( GA) in search of the most efficient design. As a procedure,a large number of initial designs were generated randomly. Subsequently,the best candidates are randomly paired and exchange their sub-parts,generating new designs of higher efficiency in an iterative manner. Despite its robustness in obtaining the optimal event timing,the GA approach fails to maintain the balance of ISI length distribution across experimental conditions and scanning sessions,which introduces potential confusions when comparing the experimental effects between different conditions. Based on the respective limitations of

关 键 词：功能性核磁共振 事件相关设计优化 概率分布函数 遗传算法 

分 类 号：B841[哲学宗教—基础心理学]