A Searchlight Factor Model Approach for Locating Shared Information in Multi-Subject fMRI Analysis

by   Hejia Zhang, et al.

There is a growing interest in joint multi-subject fMRI analysis. The challenge of such analysis comes from inherent anatomical and functional variability across subjects. One approach to resolving this is a shared response factor model. This assumes a shared and time synchronized stimulus across subjects. Such a model can often identify shared information, but it may not be able to pinpoint with high resolution the spatial location of this information. In this work, we examine a searchlight based shared response model to identify shared information in small contiguous regions (searchlights) across the whole brain. Validation using classification tasks demonstrates that we can pinpoint informative local regions.



There are no comments yet.


page 3

page 4


A Convolutional Autoencoder for Multi-Subject fMRI Data Aggregation

Finding the most effective way to aggregate multi-subject fMRI data is a...

Time-Resolved fMRI Shared Response Model using Gaussian Process Factor Analysis

Multi-subject fMRI studies are challenging due to the high variability o...

Enabling Factor Analysis on Thousand-Subject Neuroimaging Datasets

The scale of functional magnetic resonance image data is rapidly increas...

A shared neural encoding model for the prediction of subject-specific fMRI response

The increasing popularity of naturalistic paradigms in fMRI (such as mov...

Fast shared response model for fMRI data

The shared response model provides a simple but effective framework toan...

Gradient Hyperalignment for multi-subject fMRI data alignment

Multi-subject fMRI data analysis is an interesting and challenging probl...

Locally-Optimized Inter-Subject Alignment of Functional Cortical Regions

Inter-subject registration of cortical areas is necessary in functional ...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

The sensitivity of the statistical data analysis of fMRI data is limited by amount of data available. Hence there is an increasing interest in gathering data from multiple sessions and/or multiple subjects. It is hoped that this will increase statistical sensitivity in testing hypotheses about human cognition [1, 2]. Here we focus on the problem of analyzing fMRI data across multiple subjects all experiencing the same time synchronized stimulus. The raw fMRI data is thus temporally aligned, but is neither anatomically nor functionally aligned across subjects. Traditional methods of addressing this via anatomical registration, [3, 4, 5], ignore the inherent functional variability across subjects. More recent approaches attempt to use anatomical and functional data to align, or register, functional structure [6, 7, 8]. A distinct approach learns a factor model that jointly models the functional responses. Such factor methods have shown promising results and are attracting increased attention, e.g., [9, 10, 11, 12, 13, 14].

To make things more precise let denote the data from subject , , where are the number of voxels in the 3D volume along the axes, and is number of time samples (TRs) in the experiment. Typically one re-arranges into a by matrix, where . If one is only interested in voxels within a given region of interest (ROI), then is simply the sub-matrix formed by restriction to those voxels. In this setting, the basic goal is to jointly factorize the matrices into a product so as to minimize some constrained or regularized objective function. Here is the number of factors, , and . The “mixing matrices” are subject specific (acknowledging the distinct functional topographies of the subjects), while the “source matrix” provides a set of shared elicited time responses. For this reason we call such models shared response models.

Form by stacking the along the spatial dimension, and similarly form . Then where is the shared response and is the model error. and are then determined by minimizing a regularized cost function of the terms in the decomposition subject to constraints on and .

As a specific example, PCA selects and to minimize . The resulting has orthonormal columns and is partitioned into sub-matrices , . Since , the sources are formed as linear combinations of potentially all voxel time courses. Similarly, using a compact SVD: , , and is the first columns of . So the columns of are linear combinations of potentially all columns of .

In general, in a shared response factor model can be formed as a linear combination of all columns in and can be formed using a linear combination of all voxel time courses. In this sense the factor model has no assurance of preserving spatial locality. Such methods are usually applied to large pre-defined ROIs, such as PMC [14] or ventral temporal cortex (VT) [13], to investigate the relationship between the ROI and a specific stimulus or cognitive state.

In this paper we specifically set out to preserve spatial locality in forming the shared response factor model. This is important in exploratory analyses where one seeks to detect and study small local regions in the brain where shared information is present. There are two possible approaches: a searchlight analysis [15, 16, 17, 18], or adding spatial regularization to the factorization cost function. Both approaches are worthy of study. Here we focus on the searchlight approach.

A searchlight uses a fixed number of neighboring voxels to conduct analysis for each voxel location, and the same analysis is conducted on all locations. This idea can be used to extend any ROI-based factor model to overlapping searchlights. We combine factor models of the SRM form and searchlight analysis to enable localized analysis in the whole brain multi-subject fMRI functional alignment. In detail, a fixed sized searchlight centering at voxel is used to scan over the whole brain. For each searchlight location, a factor model is used to functionally align across subjects, and an analysis is performed based on the results of the alignment. Statistics from the analysis (e.g., classification accuracy) is assigned to the center voxel of that specific searchlight. We report the accuracy of each searchlight on a given classification task. This helps identify locations in the brain in which information is shared across subjects for a specific cognitive task.

For multi-subject neuroscience datasets and experiments, we provide an effective method for locating where the shared information is over the whole brain while keeping the quality of the found shared information. So our method can serve as a first step in multi-subject fMRI analysis to help identify regions that worth further investigation in a neuroscience experiment. In some factor models, the number of latent factors can be pre-specified. In this case, we record the k value that gives the best analysis result for each searchlight. We report both the accuracy and the best value on a brain map as a proxy for the presence and richness of a shared cognitive state across subjects. We also developed two new variants of the ICA and group ICA factor models. These variants show good performance in the functional alignment task.

1:Data matrices , number of factors , convergence threshold , max iteration , number of subjects
2: initialization with random orthonormal columns
3:for  in to  do
4:      is pseudo-inverse
5:     for  in to  do
8:     converged
9:     for  in to  do
10:         if max then
11:              converged               
12:     if converged then break      
Algorithm 1 Shared Response ICA (SR-ICA)
1:Data matrices , number of factors , number of subjects
2:for  in to  do
3:      First PCA with components
5: Second PCA with components
6: ICA with components
Algorithm 2 Shared Response Group ICA (SR-GICA)

2 Some Specific Factor Models

Examples of (probabilistic and deterministic) factor methods that have been used in multi-subject fMRI analysis include PCA [9], ICA [10, 11], Group ICA [12], hyperalignment (HA) [13], and shared response model (SRM) [14, 19].

The desirable factor models to be combined with searchlight analysis in multi-subject functional alignment should have the following properties: 1) has an adjustable number of factors ; 2) shows good performance in large area multi-subject functional alignment. Below, we discuss and compare several candidates using a consistent framework.

PCA: Standard PCA is a deterministic model. We have already outlined how to obtain the and in this case.

SRM [14]: The shared response model in [14] minimizes


subject to . This is done in a graphical model framework that provides Bayesian regularization. This SRM can be considered as a variant of pPCA in a multi-source setting [20, 14]. The shared response of some held out data for subject can be computed as .

Hyperalignment (HA) [13]: HA is a deterministic model that learns orthogonal to minimize (1). It does not have a selectable factor dimension . For , the non-probabilistic version of the SRM in [14] yields HA. So HA can be considered a special case of the SRM in [14]. A searchlight application of HA has been explored in [17]. For these reasons we do not explore HA further here.

Independent Component Analysis (ICA):

ICA learns statistically independent signals as measured by kurtosis or negentropy

[10, 11]. We use the FastICA algorithm, an efficient probabilistic method [10] optimizing negentropy of the shared response . This is formulated as


where is a nonquadratic function, e.g. , and

is a standard normal random variable. This yields

. is then partitioned into sub-matrices , .

SR-ICA: We also study a new algorithm we call shared response ICA (SR-ICA). Motivated by the framework in [14], we modify the FastICA algorithm. In SR-ICA, the block structure of the subject data is preserved by spatial concatenation in both and . The key difference is that instead of learning a joint matrix , we iteratively learn to ensure block-wise structure in . This is summarized in Algorithm 1111We acknowledge the help of Jacob Simon in coding SR-ICA.. Here we follow the convention of working with unmixing matrices instead of . The function is the derivative of in (2) [21].

Group ICA (GICA): Group ICA, an algorithm for making group inferences, uses two applications of PCA and an application of ICA [12]. The original algorithm first performs subject specific PCA along the temporal dimension. Then the projected data matrices for all subjects are concatenated to form a joint data matrix. A second PCA is then performed on the joint data matrix. Lastly, an ICA is performed on the projected data matrix after the second PCA. We apply GICA along the spatial dimension to learn a low dimensional shared response space. See Algorithm 2.

Dataset TRs (s/TR)
audiobook (narrated story) [22] 449(2)
sherlock-movie (audio-visual movie) [23] 1976(2)
sherlock-recall (movie free-recall) [23] 4371009(2)
Table 1: fMRI datasets used in the experiments. All datasets are whole brain (WB) in MNI [5]. We use 9 subjects in each dataset.
Figure 1: Top: Accuracy and brain maps for time segment matching using - dataset. EVC: early visual cortex.

3 Experiments and Results

Three datasets (Table 1), collected using a 3T Siemens Skyra scanner with different preprocessing pipelines, are used to test and compare the searchlight factor models. The sherlock-movie dataset was collected while the subjects watched a minute BBC episode ”Sherlock”. The sherlock-recall dataset was collected while the same subjects verbally reiterated the ”Sherlock” episode without any hints. The audiobook dataset was collected while a different set of subjects listened to a minute narrated story.

In all experiments, we use searchlights on data down-sampled by 2. For each searchlight, we try and report the highest accuracy and the corresponding value. To ensure a fair comparison, for SR-GICA, we set to be the number of voxels in the searchlight and . We tried other values, but this resulted in less accuracy. Note that there are relatively few voxels per searchlight to begin with. The accuracies are computed based on the projected shared response of held out data using learned subject-specific maps .

Time Segment Matching: This experiment is designed to test if the shared response we learned can be generalized to new data. That is, what is the quality of the shared information extracted. We use the audiobook and sherlock-movie datasets. The fMRI data are split into two halves along the temporal axis, one for training and the other for testing, and the roles reversed and the results averaged. In this experiment, we first use training data of all 9 subjects for learning the shared response. Then, a random 9 TR time segment from the testing subject (1 of the 9 subjects), called test segment, projected to the shared response space. The other 8 subjects’ testing data is projected to the shared response space and averaged. We then locate this time segment by maximizing Pearson correlation between the average response and response from the test segment. Segments overlapping with the test segment are excluded in matching. Assuming independent time segments, chance accuracy is for audiobook and for sherlock-movie. The accuracy and k brain maps for different searchlight factor models are shown in in Fig. 1 and Fig. 2. Note that we threshold the accuracy to give a more clear visualization of the most informative area. We also compute a single number accuracy by aggregating the local shared response from all searchlights. This accuracy is compared with accuracy from whole brain factor models with features ( for SR-GICA). The results are shown in Fig. 4.

Figure 2: Accuracy and k brain maps for time segment matching experiment using audiobook dataset. EAC: early auditory cortex.

Scene Recall matching: This experiment is designed to test if brain functional patterns are similar when subjects are recalling the same scene. We use the sherlock-movie for training and sherlock-recall for testing. In sherlock-recall, TRs collected when the subject was recalling the same scene are averaged and projected to the shared response space using

learned from the training data. The projected recall data along with the corresponding scene labels are used to train a SVM classifier. The projected recall data from a testing subject is used to test the classifier. Chance accuracy is

. The accuracy and k brain maps for a subset of models are shown in Fig. 3. Accuracies for searchlight factor models and whole brain factor models are computed the same way as time segment matching experiment and are shown in Fig. 4.

Figure 3: Accuracy and k brain maps for scene recall matching experiment using sherlock-recall dataset. PMC: posterior medial cortex.
Figure 4: Top and Middle: Time segment matching accuracies (top: audiobook, middle: sherlock-movie). Bottom: Scene recall matching accuracy (sherlock-movie).

4 Discussion and Conclusion

We have investigated how well various factor models can locate informative regions in a searchlight based analysis of multi-subject fMRI data. This approach highlights local brain regions that are most informative of the cognitive state of interest using both accuracy and brain maps. Early auditory cortex (EAC) and early visual cortex (EVC) are the most informative regions in time segment matching experiment for audiobook and sherlock-movie dataset, respectively. This matches the type of stimulus in these datasets. Scene recall is a more complex task. In this case a higher level cognitive region, PMC, is more informative. The results demonstrate that the approach can effectively locate meaningful informative local regions. In some neuroscience experiments, it is not clear which regions will be most relevant to the stimulus and/or task. The searchlight factor model approach helps locate regions worthy of further exploration. Moreover, since the searchlight approach preserves spatial locality, we expected the overall accuracy to drop as a consequence of the searchlight constraint. In fact, as shown in Fig. 4, the overall accuracy does not drop in most cases, and sometimes even significantly increases. The brain maps also help reveal the effectiveness of the various factor models. For example, consider the brain maps of SR-ICA and SRM. While the accuracy maps of these methods are very close to each other, SR-ICA uses a smaller to achieve this accuracy. This suggests that each factor in SR-ICA has an improved representation capability. Overall the SR-ICA is showing strong performance across the three experiments.


  • [1] Karl J Friston, Andrew P Holmes, CJ Price, C Büchel, and KJ Worsley, “Multisubject fmri studies and conjunction analyses,” Neuroimage, vol. 10, no. 4, 1999.
  • [2] Jeanette A Mumford and Thomas E Nichols, “Modeling and inference of multisubject fmri data,” IEEE Engineering in Medicine and Biology Magazine, vol. 25, no. 2, 2006.
  • [3] Jean Talairach and Pierre Tournoux, “Co-planar stereotaxic atlas of the human brain. 3-dimensional proportional system: an approach to cerebral imaging,” 1988.
  • [4] Bruce Fischl, Martin I Sereno, Roger BH Tootell, Anders M Dale, et al., “High-resolution intersubject averaging and a coordinate system for the cortical surface,” Human brain mapping, vol. 8, no. 4, 1999.
  • [5] John Mazziotta et al., “A probabilistic atlas and reference system for the human brain: International consortium for brain mapping (ICBM),” Philosophical Transactions of the Royal Society of London B: Biological Sciences, vol. 356, no. 1412, 2001.
  • [6] Bryan Conroy, Ben Singer, James Haxby, and Peter J Ramadge, “fmri-based inter-subject cortical alignment using functional connectivity,” in Advances in neural information processing systems, 2009.
  • [7] Mert R Sabuncu, Benjamin D Singer, Bryan Conroy, Ronald E Bryan, Peter J Ramadge, and James V Haxby, “Function-based intersubject alignment of human cortical anatomy,” Cerebral Cortex, vol. 20, no. 1, 2010.
  • [8] Bryan R Conroy, Benjamin D Singer, J Swaroop Guntupalli, Peter J Ramadge, and James V Haxby, “Inter-subject alignment of human cortical anatomy using functional connectivity,” NeuroImage, vol. 81, 2013.
  • [9] Hervé Abdi and Lynne J. Williams,

    Principal component analysis,”

    Wiley Interdisciplinary Reviews: Computational Statistics, vol. 2, no. 4, 2010.
  • [10] Aapo Hyvärinen, Juha Karhunen, and Erkki Oja, Independent component analysis, vol. 46, John Wiley & Sons, 2004.
  • [11] Jong-Hwan Lee, Te-Won Lee, Ferenc A Jolesz, and Seung-Schik Yoo,

    “Independent vector analysis (iva): multivariate approach for fmri group study,”

    Neuroimage, vol. 40, no. 1, 2008.
  • [12] VD Calhoun, T Adali, GD Pearlson, and JJ Pekar, “A method for making group inferences from functional mri data using independent component analysis,” Human brain mapping, vol. 14, no. 3, 2001.
  • [13] James V Haxby, J Swaroop Guntupalli, Andrew C Connolly, Yaroslav O Halchenko, Bryan R Conroy, M Ida Gobbini, Michael Hanke, and Peter J Ramadge, “A common, high-dimensional model of the representational space in human ventral temporal cortex,” Neuron, vol. 72, no. 2, 2011.
  • [14] Po-Hsuan Chen, Janice Chen, Yaara Yeshurun, Uri Hasson, James Haxby, and Peter J Ramadge, “A reduced-dimension fmri shared response model,” in Advances in Neural Information Processing Systems, 2015.
  • [15] Nikolaus Kriegeskorte, Rainer Goebel, and Peter Bandettini, “Information-based functional brain mapping,” Proceedings of the National Academy of Sciences of the United States of America, vol. 103, no. 10, 2006.
  • [16] Joset A Etzel, Jeffrey M Zacks, and Todd S Braver, “Searchlight analysis: promise, pitfalls, and potential,” Neuroimage, vol. 78, 2013.
  • [17] J Swaroop Guntupalli, Michael Hanke, Yaroslav O Halchenko, Andrew C Connolly, Peter J Ramadge, and James V Haxby, “A model of representational spaces in human cortex,” Cerebral Cortex, 2016.
  • [18] Po-Hsuan Chen, Xia Zhu, Hejia Zhang, Javier S Turek, Janice Chen, Theodore L Willke, Uri Hasson, and Peter J Ramadge, “A convolutional autoencoder for multi-subject fmri data aggregation,” arXiv preprint arXiv:1608.04846, 2016.
  • [19] Michael J Anderson, Mihai Capota, Javier S Turek, Xia Zhu, Theodore L Willke, Yida Wang, Po-Hsuan Chen, Jeremy R Manning, Peter J Ramadge, and Kenneth A Norman, “Enabling factor analysis on thousand-subject neuroimaging datasets,” arXiv preprint arXiv:1608.04647, 2016.
  • [20] Jong-Hoon Ahn and Jong-Hoon Oh, “A constrained em algorithm for principal component analysis,” Neural Computation, vol. 15, no. 1, 2003.
  • [21] Aapo Hyvärinen and Erkki Oja, “Independent component analysis: algorithms and applications,” Neural networks, vol. 13, no. 4, 2000.
  • [22] Y. Yeshurun et al., “How does the brain represent different ways of understanding the same story?,” Society for Neuroscience Abstracts, 2014.
  • [23] Janice Chen, Yuan Chang Leong, Kenneth A Norman, and Uri Hasson, “Shared experience, shared memory: a common structure for brain activity during naturalistic recall,” bioRxiv, 2016.