Neuroinformatics

, Volume 5, Issue 4, pp 223–234 | Cite as

REX: Response Exploration for Neuroimaging Datasets

Article

Abstract

Neuroimaging technologies produce large and complex datasets. The challenge of comprehensively analysing the recorded dynamics remains an important field of research. The whole-brain linear modelling of hypothesised response dynamics and experimental effects must utilise simple basis sets, which may not detect unexpected or complex signal effects. These unmodelled effects can influence statistical mapping results, and provide important additional clues to the underlying neural dynamics. They can be detected via exploration of the raw signal, however this can be difficult. Specialised visualisation tools are required to manage the huge number of voxels, events and scans. Many effects can be occluded by noise in individual voxel time-series. This paper describes a visualisation framework developed for the assessment of entire neuroimaging datasets.While currently available tools tend to be tied to a specific model of experimental effects, this framework includes a novel metadata schema that enables the rapid selection and processing of responses based on easily-adjusted classifications of scans, brain regions, and events. Flexible event-related averaging and process pipelining capabilities enable users to investigate the effects of preprocessing algorithms and to visualise power spectra and other transformations of the data. The framework has been implemented as a MATLAB package, REX (Response Exploration), which has been utilised within our lab and is now publicly available for download. Its interface enables the real-time control of data selection and processing, for very rapid visualisation. The concepts outlined in this paper have general applicability, and could provide significant further functionality to neuroimaging databasing and process pipeline environments.

Keywords

REX Exploration fMRI Response Analysis Metadata ROI Visualisation GLM Neuroimaging Region of interest 

Introduction

Standard neuroimaging analysis tools fit general linear models (GLMs) to voxel time-series and map regions displaying statistically significant experimental effects. The approach is statistically sophisticated and highly flexible. Complex multi-factorial and mixed-effects experimental designs can be assessed, using procedures that account for autocorrelated noise and spatially correlated signal. However the need for parsimony, and the limited number of trials available for modelling, motivate the use of simple signal models. Typically, a small set of uncorrelated regressors are used to model voxel signal responses, and the effect of experimental factors on the resulting response parameters is assumed to be linear. Unexpected or complex signal effects are difficult to assess in this analysis. For example, deactivations, response habituation, unmodelled responses, and the effects of motion and preprocessing choices can all affect statistical mapping results, yet easily go undetected (LaConte et al. 2003; Razavi et al. 2003; Rekkas et al. 2005). BOLD responses can show rich temporal dynamics that may exhibit interesting changes across experimental conditions, however they are difficult to incorporate into simple models and tend to receive little attention in routine analysis (Duff et al. 2007a; Fox et al. 2005; Harms et al. 2005).

Exploratory visualisation of datasets can detect many of these effects, providing insight into underlying neural activity and motivating improved data modelling and experimental design. Due to the size and complexity of neuroimaging datasets, development of methods for efficient and systematic data exploration is an important area of research. Luo and Nichols (2003) focused on the detection of exceptions to linear modelling assumptions, using diagnostic statistics and visualisations of experimental data. Their tool, SPMd, integrates this approach into the SPM package (www.fil.ion.ucl.ac.uk/spm/ext/#SPMd), and incorporates key tenants of exploratory data analysis (EDA; Tukey 1977). Interactive windows of linked orthogonal slice viewers enable the visualisation of maps of diagnostic statistics and model and voxel data, with further linked windows displaying outlier and motion time-series and other characterisations of the modelling. SPMd can be used to assess the modelling of individual scanning sessions as well as the modelling of experimental effects within a series of lower-level parameter maps. It is useful for identifying a range of artefacts and model deficiencies at different levels of the experimental hierarchy.

However, high levels of signal noise mean many systematic effects, such as response habituation, complex response shapes and the effects of motion, will not cause detectable deviations from modelling assumptions and may not be evident in individual voxel time-series. Detection of such effects may require the assessment of many responses using data averaging and other processing. A number of tools facilitate this type of exploration. The GLM package Brain Voyager (www.brainvoyager.com) enables the visualisation of event-averaged responses associated with an analysis. Users define a data-file specifying the scans and event-classes to be combined for averaging and other data extraction parameters. This data is then used to generate plots of averaged responses as the user browses activation maps, providing useful insight into the signal that has produced the activations. Gadde et al. (2004) have developed a set of command-line tools (BXH-tools) for the extraction and processing of event responses that match specified characteristics. Data selection is independent of a specific linear model of experimental effects, being based on experimental meta-data defined within the BXH or XCEDE schemas (Keator et al. 2006, http://www.nbirn.net/downloads/bxh_tools/). This enables very flexible selection of data from across an experiment. However these tools have not been integrated into an exploratory graphical user interface.

While many processing steps may be utilised in the detection and characterisation of signal features, such as data interpolation, prewhitening or the calculation of signal variance, current exploration tools tend to be limited in their processing options. Flexible, real-time control over processing would enable more interactive investigation of signal effects and would be useful for the assessment of preprocessing algorithms. One approach is to utilise standardised processing mod ules that can be linked to define an executable processing stream with a minimum of user input. A number of “process pipelining” environments have been developed for general neuroimaging applications (Fissell et al. 2003; Rex et al. 2003). Wrappers are used to integrate third-party processing algorithms as processing modules. The processing streams are defined in user-friendly graphical environments, and the allocation of processing resources managed by the software. These tools have not been widely utilised for functional data analysis. While preprocessing strategies might be explored using pipeline tools, integrated fMRI analysis packages effectively automate specific analysis streams and are purpose built to enable the rapid setup and estimation of the complex signal models. For fMRI data exploration, non-specialised pipeline environments may not provide the necessary data management or interactivity.

There is a need for dedicated software tools for flexible and interactive exploration of experimental datasets. This paper discusses design and implementation issues faced in the development of such applications, and describes a new exploration tool, REX, that provides users with extensive, real-time control over the extraction, processing and visualisation of signal from across neuroimaging datasets. The tool facilitates systematic exploration of raw and averaged event-related responses from regions-of-interest (ROIs), and rapid assessment of the impact of any experimental factor or processing step on these data. Users are able to define processing modules that can be applied at different stages of the overall processing stream. Data organisation is easily defined, and can be automatically imported from FSL Feat or SPM linear modelling design files (www.fmrib.ox.ac.uk/fsl/feat5/ and www.fil.ion.ucl.ac.uk/spm/). This information is automatically utilised to customise the exploration interface. Implemented in MATLAB, REX has been widely utilised within our laboratory. It is easy to use and can provide valuable insight into properties of the signal and noise underlying analysis results in a wide range of experiments.

Methods

Overview

The goal of the work was to design a tool providing flexible control over data extraction, processing and visualisation from within a single window. We focused on the visualisation of event-averaged data, using ROIs. Averaged ROI data can provide valuable insight into the underlying signal dynamics, and a set of ROIs provides an effective basis around which a thorough but efficient assessment of a dataset can be structured. Currently available tools do not enable simultaneous assessment of ROIs and can be slow and awkward when used for extensive exploration.

REX’s basic workflow is straightforward. The organisation of the dataset is first specified in a setup interface. Much of the information can be automatically determined from SPM or Feat analysis files. ROIs are selected and relevant voxel data can be extracted into memory. The resulting data can be saved for ongoing use and may be updated at any time. The exploration interface consists of a control panel, a data visualisation window, and a data table (Fig. 4). In the control panel, users can adjust data selection, data processing and visualisation settings. A “Plot” button initiates data processing and output is displayed in a selected row of axes within the visualisation window. User defined “response measures” are tabulated in the data panel below. This output can be saved for further analysis in MATLAB or text format.

Key aspects of development included designing a schema for the experimental meta-data, designing the process pipelining system, and organising the look and feel of the user interface. The meta-data schema needed to accommodate a wide range of experimental designs, yet maintain a simple structure that could be both easily edited by users and automatically integrated into data selection and processing menus in the exploration interface. Similarly, the process pipelining system had to be capable of implementing a wide range of algorithms, however needed a structured organisation for it to be intuitively controlled via a simple interface.

Data Indexing and Selection

Neuroimaging experimental datasets have a complex structure, typically consisting of numerous scanning runs recorded across one or more scanning sessions for each subject, and multiple subject groups. Scans may be truncated or contaminated by artefact. The data is spatially complex, consisting of functional regions of varying sizes whose location varies from subject to subject. It is also temporally complex, typically recording activity associated with randomised switches between different task conditions, with a background of other signal fluctuations and hardware noise. In addition to the imaging data, there is often auxiliary time-series data such as physiological measures or task performance ratings that are important to the analysis. Carefully designed software is essential to effectively manage such datasets.

An exploration tool must enable users to organise data selection and processing based on specific features of a dataset’s structure. Users also need be able to easily edit this organisation, for example, adding new scan data or defining new experimental factors such as new subject or task classifications. These requirements demand a simple, flexible organisation of the experimental metadata. We identified three core data-elements—scanning runs, time-series data, and scan events—that can be used as the basic elements upon which the organisation of a neuroimaging experiment can be defined. Scanning runs refer to any experimental period in which data is recorded. Scan events are periods within these runs, defined by onset and duration. Time-series data refers to any data recorded during a scanning run. Time-series data can include two classes, voxel-data, which has a defined spatial location, and additional “covariates”, that might include ongoing physiological measures, head-motion estimates or reaction-time measures.

Experimental organisation can be specified by indexing these core elements to define different clas sifications of the data. For example, subjects can be identified by associating with each scanning run an index specifying subject ID. This simple organisation ensures a high level of flexibility, and is simple to integrate into the exploration interface. In addition to identifying subjects, scanning run classifications could also indicate the age of the subject or whether they are members of a patient group. Event indexing could specify task types, onset times, or performance scores. It might also be used to flag responses, or other periods of the experiment, thought to be contaminated by artefacts. Voxel-data time-series classifications can identify ROIs, image slices, or functional networks (Fig. 1).
Fig. 1

Schematic representation of the meta-data schema employed in REX. Three core data elements are indexed—scanning runs, events, and data time-series. Classifications of these elements—user-defined, or automatically generated from FSL or SPM design files—are automatically incorporated into data selection, grouping and processing menus of the exploration interface. Scan classifications can specify which subject, session or subject-group an individual scan belongs to. Time-course indexing can specify voxels belonging to regions of interest (ROIs), and can tag ICA or motion time-course data. Event classifications can indicate various features of responses, such as task-type or response scores

In REX, the primary data-classes are stored in separate, linked arrays. Most experimental data, such as image or event-file locations, is stored within the elements of one of these data classes. Data indexing is achieved through user-defined entries within elements of any of the classes. With a focus on ROIs, the indexing of voxel time-series has been streamlined. All voxels composing a ROI are stored and indexed as a single data element, with covariates stored in a separate array. Time-series data is pre-extracted from disc during the setup procedure to optimise the speed of data processing.

The REX data setup interface directly reflects its meta-data organisation. Scans, events, covariates and ROIs are defined in separate windows. A range of automated functionality assists setup (Table 1). REX can determine scan and event data elements automatically from SPM or FSL analysis design files. Events are classified based on the regressors from which they were extracted. Continuous regressors and head motion estimates are stored as covariates. Scans can be automatically classified based on the experiment’s directory structure, and events classified according on onset times and duration. Data-elements can be manually defined and edited en-masse.
Table 1

Automated data extraction and indexing

Data element

Determined from

Indexed according to

Sessions

Image files

Directory hierarchy

SPM .mat files

Group-level model

FSL .fsl files

Events

Discrete GLM regressors

Regressor name

Onset time

Event duration

Covariates

Motion parameters

Axis

Continuous GLM regressors

Regressor name

Voxels/ROIs

MNI or native space mask

File name

SPM VOI file

Some care is required in the indexing of events. For event-related averaging, the conditions immediately prior to the onset of a selected event-type may vary. Analyses of different transitions can sometimes reveal interesting differences in transient responses reflecting the cognitive demands of specific task-switches. Different transitions can be easily specified and investigated using REX.

ROIs are specified using 3D mask images or SPM volume-of-interest files, either defined separately in each session’s native space, or in normalised space. REX uses SPM or FSL functions to invert the spatial normalisation, enabling the extraction of voxel data from native space. This permits the direct use of slice timing information in processing, enables accurate assessment of ROI homogeneity, and reduces memory usage.

Once the experimental organisation has been fully defined, and the voxel data extracted, it can be loaded into the exploration interface (Fig. 4). Data selection is performed via a series of menus listing the classes defined within different classifications of data. Scans are selected for processing through adjusting a pool of scans by selecting or deselecting different scan classes. For example, the user may select the patient group, from the “group” scan classification menu, then deselect subject number four in the “subject” menu. Events are selected from a single menu listing all defined event classes. Regions of interest and covariates are chosen via another menu. If multiple ROIs or covariates are selected they are processed simultaneously and plotted separately across the screen.

Data Processing

Extraction and assessment of event-related response data requires a significant amount of processing. Prior to event extraction, baseline drifts and fluctuations should be estimated and removed. The data might also be converted to percentage signal change, filtered or whitened. Event extraction will involve cutting out sections of the scan time-series corresponding to experimental events of interest, and temporally aligning this data. A window of at least 10 s surrounding an event needs to be extracted to ensure all event-related signal effects are captured. Further processing follows extraction. Resampling of the response data may be necessary to account for jittered sampling of events. For ROI analysis, a representative response needs to be calculated from the voxels within the ROI. This can be achieved by averaging the voxel data, or by calculating principal components. After this, further processing may be applied, such as model fitting or Fourier transformation, or the data may be averaged across one or more levels of the experimental hierarchy.

Processing requirements vary considerably depending on the aims of the analysis. A process pipelining system is ideal way of providing the required processing flexibility in an analysis tool. To be effective in an exploration context, we designed a semi-structured system, based around a core processing sequence that reflects the different stages of event-related processing (Fig. 2). Separate processing modules are defined for processing prior to and following response extraction, and following event grouping. A single menu provides a list of all available processing modules. The order in which they are selected defines the order they are applied to the data. With this simple system, the pipeline can be set up very rapidly. Event extraction and averaging are hard-coded components of the processing stream, that can be controlled via separate sections of the exploration interface. The grouping and averaging of responses after they are processed is also hard-coded. Both event and scan classifications can identify interesting experimental factors, so both are listed in the data averaging menu, along with options of having no averaging, of applying only the ROI averaging (plotting individual responses), or of averaging across all responses. Selecting multiple factors results in data being organised into appropriate subgroups. This structured organisation allows the entire processing stream to be quickly and intuitively specified, with basic response visualisation requiring minimal interaction with pipeline modules.
Fig. 2

The REX processing stream. A series of user defined data processing steps can be defined prior to and following ROI averaging and event extraction, and following event grouping. Multiple ‘response-measures’ can be defined, which are averaged in parallel with the main processing stream. Both time-course data and measures can be visualised

Processing modules consist of single line MATLAB commands. The simple language means users with limited programming experience should be able to quickly learn to define their own modules (Fig. 3). As modules may modify both signal and time variables, and may call any available MATLAB function, complex data transformations can be implemented. ASL or motion measurements can be differenced, data resampled, and Fourier or wavelet transforms can be applied. Model fitting modules may be defined and the results overlaid upon the raw responses. Sets of new processing modules can be saved for use across various projects. As processing can cause some minor delays in the interface, a number of stages of caching have been implemented, storing processed data until the processing settings are altered.
Fig. 3

Definition of response measures. Any MATLAB function can be called in from a response measure or processing modules

In addition to the primary processing stream, additional “response measures” can be employed to provide single-value summaries of the response data. Response measures are calculated immediately prior to the averaging of response groups. They are averaged in the same manner as the main processing stream, with the results tabulated below the visualisation screen. These measures enable the rapid assessment of the sensitivity of basic aspects of responses to experimental factors. Predefined response measures include mean signal, signal standard deviation, a voxel count, and diagnostic statistics (Table 2). New response measures can be defined in the same manner as processing modules, with users able to specify thresholds for the systematic detection of outliers.
Table 2

Examples of basic processing and response-measure modules

 

Function

Acting on..

Notes

Processing Module

Remove mean

Scans

 

Detrend

Scans

Does not remove mean

High pass filter

Scans

Butterworth > 0.005 Hz

Percentage signal change

Scans

Based on mean scan signal

Low pass filter

Scans

Butterworth < 0.1 Hz

Difference time-series

Scans

For assessment of avg. motion

Sub. baseline prior to event

Events

Mean of signal prior to onset

Calculate power spectrum

Group

To assess average spectral props

Response Measures

Number of voxels

Scans

 

Signal mean

Events

Period 8 s post-onset to offset

Signal standard deviation

Events

Period 8 s post-onset to offset

Onset

Events

Indicates onset time

Outlier count

Events

No. time points > 3 SD from mean

Onset/offset signal

Events

 

Statistics

GLM analyses assess a small number of narrowly defined hypothesised effects through spatial mapping. Exploratory analyses, on the other hand, aim to identify unexpected, dynamics, employing interactive visualisation to enable the identification of effects. With no well defined prior hypothesis, statistical tests are not central. Nevertheless, statistical tests can play a useful role in directing exploration. REX provides the option to automatically apply a two-way ANOVA to the response data, with factors of time and the chosen response groupings (e.g. response type or subject group). A plot is flagged if there are any significant effects. There may be an effect of time (e.g. any consistent response time course over all groups), group (e.g. different average levels of signal across the response periods across groups), or a significant interaction effect (e.g. a significant difference in response shape between groups). In addition, ANOVAs are calculated for all selected response measures, testing for between-group differences. These results are intended to be used as pointers to interesting effects, useful for directing the exploration and identification of potential confounds. For example, by selecting the voxel count measure, which counts the number of voxels in the ROI for each scan, the user will be alerted when there are significant differences in ROI size across groups.

Visualisation and Interaction

The output of the data processing is displayed in the visualisation window (see Fig. 4). This window can be split, twice, both horizontally and vertically, so that it may contain between one and nine plot axes. The processed time-series are displayed in the selected row, with the results for each ROI tiled across the axes in this row. Any number of ROIs can be cut scrolled through across the horizontal plane. Users can also scroll downwards to reveal free plot rows, so are able to maintain a large number of data-plots in memory. Plots may be overlaid upon each other. Below the plot window is a summary table for a selected plot. Users can switch between a view of individual event measures and a summary table, which tabulates the mean and standard deviation of statistics for each data group. Individual responses can be selected, and overlaid upon the average data. In the former view, any response-measure exhibiting significant effects according to the ANOVA statistics, or individual measures exceeding predefined thresholds, are flagged. All the data associated with a selected plot, including the error bars and event statistics, are stored in a single data structure, which can be saved for detailed analysis in MATLAB or text format. The user can also choose to plot the response measures as a pseudo time-series across the one or more factors according to which the data was grouped.
Fig. 4

The REX interface, showing an investigation of a small experimental dataset recording a motor-task. At top left, the “Data Selection” panel shows that scans are classified into Subject and Session groups. Subject 11 has been deselected. Cingulate, SMA and right motor cortex ROIs have been selected. Each scan consists of two 2-min task periods. The Event menu lists various event classes associated with these task periods. The “full” class corresponds to a default event consisting of the full duration of each scan. The selected events, A2–A4, select the first three task-periods of each scanning session. These have been chosen to determine if there are any apparent changes in responses in the early stages of each session. The highlighting of “Event Types” from the response averaging menu in the “Data Processing” panel means responses from each event class will be grouped and averaged separately. A standard processing stream has been selected using the processing menu, and mean signal and standard deviation response measures have been selected. Standard visualisation options have been selected within the “Data Visualisation” panel. The visualisation window has been split into a 3 ×3 matrix to compare the different ROIs and simultaneously assess a number of plot types. The currently specified analysis has been plotted in the top row, with each set of axes plotting data from one of the ROIs. Each ROI exhibits a distinct response shape. The initial task-period has been highlighted in blue. The second row of the visualisation window shows plots resulting from the addition of a processing step calculating the power spectrum of the same segment of the response data. The bottom row shows change in the average mean signal response measure over the eight task blocks in all scanning sessions (A2–A9). This style of plot can be selected in the “Data Visualisation” panel. Cingulate and SMA mean responses appear to reduce over time. Below this, the data window tabulates data associated with the selected top-left set of axes

Results

Interactive, ROI-based exploration has proven a valuable and efficient method for assessing a wide range of signal dynamics within experimental datasets. Exploration using REX regularly identifies previously unexpected signal features and artefacts, and has prompted a number of comprehensive studies of specific features. Its most straight-forward use is for the visualisation of responses and assessment of the effects of key experimental factors. Duff et al. (2007a) used REX to explore task-related responses in a motor learning experiment involving twelve subjects scanned three times over a one-month training period. Responses across the brain were found to show a variety of different temporal profiles, that were consistent across subjects. This finding motivated a thorough GLM analysis utilising an extended, multi-component “OSORU” basis set which modelled different response features including spikes at task onset and offset, and undershoots (Harms and Melcher 2003). The analysis characterised the spatial distribution of response transients and tracked changes in response shapes over the course of the study. REX was useful during this analysis, enabling the visualisation of the average responses corresponding to shifts in multiple model parameters. New REX processing modules were defined to calculate the OSORU response model fits to ROI data, enabling model fits to be overlaid upon responses, and the generation of time-plots indicating how different model parameters changed over the course of the scanning sessions.

Using REX, Cole et al. (2006) identified extended pain-related responses in Alzheimer’s patients, compared to aged control subjects. This effect is not easily detected or modelled with a GLM, yet can be of significant interest. In this case, it indicates that the Alzheimer’s patients’ attention to the noxious stimuli is prolonged. In an fMRI study of the urge to cough, REX was used to extract ROI data for further statistical analysis (Mazzone et al. 2007). Exploratory visualisation during GLM analysis revealed an unexpected response to a visual cue, and some model timing errors, which both required amendments to the basic model. Further assessment revealed unexpected deactivations in some regions during a condition involving innocuous stimuli (M. Farrell, private communication). A processing module can be used to calculate the Fourier transform of ROI data, enabling BOLD signal fluctuations to be assessed. Duff et al. (2007b) identified unexpected changes in signal fluctuations across a range of frequencies over the course of standard fMRI sessions. REX has been used to identify differences in the shape of ASL and BOLD signals, and for the processing and extraction of data from ROIs for the investigation of the balloon model (Bowala et al. 2005; Johnston et al. 2006).

These results, along with a variety of other recent reports assessing complex features of dynamics (e.g. Fox et al. 2005; Harms et al. 2005; Zhang et al. 2006), indicate the value of thorough data exploration. Such features can influence standard response and connectivity analyses, affecting statistical power, and the reliability and interpretability of results (Razavi et al. 2003). For example, the observed strength of transient dynamics in the motor system suggests that standard modelling outcomes will be sensitive to task duration.

The ability to quickly explore signal across an experimental dataset improves users’ intuitions for BOLD dynamics, and the noise, artefacts, and experimental factors that affect this signal. The straightforward interpretability of the averaged responses, and REX’s ability to rapidly visualise features of interest, makes it a useful tool with which to explain functional imaging results to collaborators and others unfamiliar with the technology. We find response visualisation encourages the investigation of non-canonical response shapes and the development of novel experimental designs and modelling strategies to account for these effects.

We find certain features of experimental design can improve the quality and interpretability of ROI signal averaging and visualisation. Adequate periods of a rest or control condition are important for determining baseline signal levels. These periods should be long enough to ensure that the response signal returns to baseline before the following response. Baseline periods at the beginning and end of each run ensure that slow signal drifts can be removed, while extended periods of baseline (e.g. > 60 s) can enable the assessment of background fluctuations and extended task-related signal effects (Duff et al. 2007a).

REX is fast to set up. If FSL or SPM design files can be utilised to load the experimental data, the REX meta-data structure can be created in 10 min for simple designs. The extraction of 20 ROIs from thirty runs of 200 MRI volumes takes approximately 2 min on a standard Pentium-class desktop machine with 2 GB of RAM. The exploration interface is also fast to use. There can be small delays in plotting when large amounts of data are processed and many ROIs are visualised simultaneously.

Visualisation of averaged responses can be achieved very quickly and with little instruction, utilising the automatically detected scan and event classes and the default processing stream. However, comprehensive analyses often require the specification of new experimental factors and the creation of additional processing modules to assess a greater range of experimental effects. An understanding of REX’s data organisation is required to incorporate additional experimental factors, while an understanding of its processing stream, as well as basic MATLAB scripting, is necessary for the assessment of different processing algorithms. An understanding of the distinct methods and aims of EDA is important to appropriately manage the flexibility provided by the tool. We have found a tendency to restrict exploration to a single level of the experimental hierarchy, and to visualise data from just a few key ROIs. This approach limits the inferences that can be made regarding the specificity of observed effects. A walk-through tutorial and user guide have been developed to assist users to effectively utilise the tool, and are available on the website.

Discussion

With complex and poorly characterised dynamics, fMRI datasets warrant detailed exploratory investigation, which can both guide and interpret hypothesis-driven modelling analysis. The results of the work described here indicate that automated exploration of entire experiments, structured around the assessment of ROI data, is practicable and can be highly effective. REX organises and automates data selection, extraction and processing, enabling rapid and flexible exploration of signal from across an experimental dataset. It can be utilised for a variety of tasks: to assess unmodelled data for artefacts, task-correlated motion, major signal effects and influential experimental factors; to interpret activations through visualisation of the underlying signal effects; to assess difficult-to-model response effects; and to extract and pre-process data for subsequent connectivity analyses and other model ling applications.

REX utilises ROIs to summarise signals from across the brain. This is a natural data reduction approach that can resolve subtle effects by combining signal from functionally homogeneous regions and removing spatially uncorrelated noise (Brett et al. 2002). It makes possible a systematic and reasonably comprehensive visualisation of signal from across the brain. However, obtaining consistent, homogeneous ROIs is not straightforward. Approaches based on anatomical structure have the advantage of being independent of the functional data. However they cannot reliably identify all functional regions, and may miss unexpectedly responsive brain regions. Reliable localisation of homogeneous regions based on functional data can be difficult on a single-subject basis. However group activation maps may not accurately localise responses for all subjects. ROI definition can also be problematic when activation is very widespread, and in experiments that generate multiple contrast maps of interest. The use of functional data may also be biased towards the detection of voxels showing signal effects matching the GLM’s response model. The functionality provided by REX helps alleviate some issues associated with the use of ROIs. It enables ROI homogeneity to be assessed through the visualisation of voxel data and the use of response measures. Its facility to rapidly visualise many ROIs aids the assessment of the spatial specificity of effects. Issues relating to ROI selection should not greatly limit the effectiveness of exploration. An ROI selection approach will be clear for many datasets and analyses. While inhomogeneous ROIs may reduce or blur signal effects, they are unlikely to produce large artefactual features. The wide range of subtle, consistent signal effects identified using REX suggests the approach is productive. In future work we plan to develop exploratory spatial mapping tools to assist the definition and exploration of ROIs.

It is important that exploration is applied appropriately at different stages of an analysis. Prior to hypothesis-driven analysis, data can be thoroughly assessed for the identification of artefactual signal effects that should be accounted for in the modelling. However, analysts must avoid visualising factors central to the experimental hypothesis, as modifications to the model influenced by these observations invalidate the analysis. An alternate approach is to assess a subset of the data in detail, using the results to generate a hypothesis that can be tested on the remainder of the dataset. Luo and Nichols (2003) employ a hierarchical approach to their model exploration and diagnosis, refining session-level modelling prior to the incorporation of these analyses into higher-level modelling of the main experimental effects. They recommend a cautious approach to the removal of artefact contaminated data, and suggest grading significant regions in the final contrast map as valid, questionable or invalid, depending on whether the model assumptions are met within those regions. Changes to response models also need to be made cautiously, based on observations made from a wide range of regions, and from an unbiased sample of the experimental dataset.

If specific signal features are consistently detected with exploratory techniques, automated approaches should be developed that optimise their detection and processing or modelling methods developed to account for their effects. Hardware induced signal spikes or slice dropouts, for example, are conducive to automatic detection and removal (e.g. the 3dDespike function in AFNI, Cox 1996). BOLD signal spikes at the onset and offset of task blocks exhibit reliable timing and may be modelled using additional regressors in the GLM framework (Harms and Melcher 2003). REX’s ability to automatically flag scans or events in which user-defined response measures exceed a specified threshold enables users to experiment with methods for detecting artefacts. If successful, these methods need be developed into preprocessing modules assessing all voxels across the brain.

Unconstrained exploration should follow (and be driven by) the hypothesis-based assessment of a dataset, providing further interpretive results and identifying unexpected effects that can guide the design of follow-up experiments. This assessment can help determine the dynamics of active regions in greater detail, and assess whether absent effects may be due to a lack of statistical power, artefacts, or poor modelling. Care needs to be taken to avoid biased assessment. An excessive focus on regions central to the cognitive theory motivating a study may mean similar effects are not noticed in other brain regions, resulting in hypotheses based on an incorrect understanding of the spatial extent of effects. One approach to avoid bias at all levels of the analysis could be to assess activity blind to anatomical locations.

The semi-structured process pipeline approach of REX is a novel design feature that facilitates the balance of flexibility and structure provided in the user interface. Further refinement of this system should continue to expand processing options, while maintaining usability. Ideally, exploration would utilise a full-scale process management system, capable of efficiently managing whole-brain processing and parallel processing of sessions. Processing modules generated during exploration could then be directly inserted into the standard analysis processing stream. Whole-brain processing will demand a wider variety of visualisation and data interaction options. A highly modular environment might also facilitate interface configurability. A more structured exploration interface, like SPMd, would be useful to guide efficient and appropriate exploration at certain stages of analysis.

The data structure utilised in REX is designed to be able to represent a wide range of experimental designs in a simple, efficient manner. Recently, large scale databasing and automated analysis efforts have motivated the design of general-purpose schema to store all experimental information within a standardised data-structure (Keator et al. 2006). A standardised meta-data format can enable the automated interchange of experimental data between software programs, and facilitate the rapid development and wide uptake of new tools. The current version of the XCEDE schema uses a five-level schema consisting of patient, subject, visit, study and session levels. It stores all standard experimental data as well as statistical results and details of the data-processing providence. REX’s schema stores data entirely at the session level. While the incorporation of other levels of structure is necessary for a full meta-data schema, the use of scan, event and time-series data elements for indexing is a clean and highly general approach to specifying a dataset’s structure. The time-series indexing implemented in REX is novel, enabling the flexible incorporation of ROIs and covariates into the schema. The ability to store the associated time series data in the meta-data file enables it to be used as a compact, easily explored summary dataset for an experiment. REX is a novel example of how a wide range of experimental meta-data can be tightly integrated into software functionality. Such an interface could be integrated into specialised neuroimaging databasing or process pipelining environments, significantly extending their functionality.

Conclusion

We have developed an exploratory analysis platform for fMRI data that automates event-related averaging with fast and flexible data selection, processing, and visualisation. REX enables comprehensive assessment of responses and other signal from across an experiment, and assessment of different experimental factors and preprocessing choices. It provides a high level of interaction with the data, often providing in valuable insight into signal properties. The tool has enabled the identification and investigation of a range of phenomena that would otherwise have gone undetected. Highly automated and modular analysis environments that support intuitive and configurable interfaces are ideal platforms for both hypothesis-driven and exploratory analysis of fMRI data. REX is a useful prototype of such an environment, with many avenues for fu ture development.

Notes

Acknowledgements

Gary Egan has been supported by a GE NHMRC Principal Research Fellowship #400317. Thanks to Leonie Carabott, Michael Farrell, Hamed Asadi and Leigh Johnston.

 

Information Sharing Statement Source code for REX is available from www.neuroimaging.org.au/nig/REX/. REX has also been registered at The Neuroimaging Informatics Tools and Resources Clearinghouse www.NITRC.org.

References

  1. Bowala, T., Scozzafava, A., Egan, G., Farrell, M., Van Meter, J., Wang, J., et al. (2005). Comparison of continuous arterial spin labeling (CASL) with blood oxygen level dependent (BOLD) contrast fMRI in functional imaging of a visually cued motor task. In Program No. 454.16.2005 Abstract Viewer and Itinerary Planner. Washington, DC: Society for Neuroscience. Online.Google Scholar
  2. Brett, M., Johnsrude, I. S., & Owen, A. M. (2002). The problem of functional localization in the human brain. Nature Reviews Neuroscience, 3(3), 243–249.PubMedCrossRefGoogle Scholar
  3. Cole, L. J., Farrell, M. J., Duff, E. P., Barber, J. B., Egan, G. F., & Gibson, S. J. (2006). Pain sensitivity and fMRI pain-related brain activity in Alzheimer’s disease. Brain, 129(Pt 11), 2957–2965.PubMedCrossRefGoogle Scholar
  4. Cox, R. W. (1996). AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Computers in Biomedical Research, 29(3), 162–173.CrossRefGoogle Scholar
  5. Duff, E., Xiong, J., Wang, B., Cunnington, R., Fox, P., & Egan, G. (2007a). Complex spatio-temporal dynamics of fMRI bold: A study of motor learning. Neuroimage, 34(1), 156–168.CrossRefGoogle Scholar
  6. Duff, E., Xiong, J., Wang, B., Cunnington, R., Fox, P., & Egan, G. (2007b). Slow, long-lasting changes in cortical blood-oxygenation fluctuations due to motor-task performance: Implications for fMRI analyses. In International Brain Research Organisation: World Congress of Neuroscience, Melbourne.Google Scholar
  7. Fissell, K., Tseytlin, E., Cunningham, D., Iyer, K., Carter, C. S., Schneider, W., et al. (2003). Fiswidgets: A graphical computing environment for neuroimaging analysis. Neuroinformatics, 1(1), 111–125.PubMedCrossRefGoogle Scholar
  8. Fox, M., Snyder, A., Barch, D., Gusnard, D., & Raichle, M. (2005). Transient BOLD responses at block transitions. NeuroImage, 28(4), 956–966.PubMedCrossRefGoogle Scholar
  9. Gadde, S., Michelich, C., & Voyvodic, J. (2004). An XML-based data access interface for image analysis and visualization software. In Proc. 10th Annual Meeting of the Organization for Human Brain Mapping. Budapest, Hungary, CD-ROM. NeuroImage 22(1).Google Scholar
  10. Harms, M., Guinan, Jr, J., Sigalovsky, I., & Melcher, J. (2005). Short-term sound temporal envelope characteristics determine multisecond time patterns of activity in human auditory cortex as shown by fMRI. Journal Neurophysiology, 93(1), 210–222.CrossRefGoogle Scholar
  11. Harms, M., & Melcher, J. (2003). Detection and quantification of a wide range of fMRI temporal responses using a physiologically-motivated basis set. Human Brain Mapping, 20(3), 168–183.PubMedCrossRefGoogle Scholar
  12. Johnston, L. A., Duff, E., & Egan, G. F. (2006). Particle filtering for nonlinear bold signal analysis. In: Larsen, R., Nielsen, M., Sporring, J. (Eds.), MICCAI (2), Vol. 4191 of Lecture Notes in Computer Science (pp. 292–299). Springer.Google Scholar
  13. Keator, D. B., Gadde, S., Grethe, J. S., Taylor, D. V., Potkin, S. G., & BIRN, F. I. R. S. T. (2006). A general XML schema and SPM toolbox for storage of neuro-imaging results and anatomical labels. Neuroinformatics, 4(2), 199–212.PubMedCrossRefGoogle Scholar
  14. LaConte, S., Anderson, J., Muley, S., Ashe, J., Frutiger, S., Rehm, K., et al. (2003). The evaluation of preprocessing choices in single-subject BOLD fMRI using NPAIRS performance metrics. NeuroImage, 18(1), 10–27.PubMedCrossRefGoogle Scholar
  15. Luo, W., & Nichols, T. (2003). Diagnosis and exploration of massively univariate neuroimaging models. NeuroImage, 19(3), 1014–1032.PubMedCrossRefGoogle Scholar
  16. Mazzone, S. B., McLennan, L., McGovern, A. E., Egan, G. F., & Farrell, M. J. (2007). Representation of capsaicin-evoked urge-to-cough in human brain using functional MRI. American Journal of Respiratory Crit Care Medicine, 176, 327–332.CrossRefGoogle Scholar
  17. Razavi, M., Grabowski, T. J., Vispoel, W. P., Monahan, P., Mehta, S., Eaton, B., et al. (2003). Model assessment and model building in fMRI. Human Brain Mapping, 20(4), 227–238.PubMedCrossRefGoogle Scholar
  18. Rekkas, P. V., Westerveld, M., Skudlarski, P., Zumer, J., Pugh, K., Spencer, D. D., et al. (2005). Neural correlates of temporal-order judgments versus those of spatial-location: Deactivation of hippocampus may facilitate spatial performance. Brain Cogn, 59(2), 103–113.PubMedCrossRefGoogle Scholar
  19. Rex, D. E., Ma, J. Q., & Toga, A. W. (2003). The LONI pipeline processing environment. NeuroImage, 19(3), 1033–1048.PubMedCrossRefGoogle Scholar
  20. Tukey, J. (1977). Exploratory data analysis. Addison-Wesley Series in Behavioral Science: Quantitative Methods. Reading, MA: Addison-Wesley.Google Scholar
  21. Zhang, H., Luo, W., & Nichols, T. (2006). Diagnosis of single-subject and group fMRI data with SPMd. Human Brain Mapping, 27(5), 442–451.PubMedCrossRefGoogle Scholar

Copyright information

© Humana Press Inc. 2007

Authors and Affiliations

  • Eugene P. Duff
    • 1
    • 2
  • Ross Cunnington
    • 3
  • Gary F. Egan
    • 1
  1. 1.The Howard Florey Institute and Centre for NeuroscienceThe University of MelbourneMelbourneAustralia
  2. 2.Department of Mathematics and StatisticsUniversity of MelbourneMelbourneAustralia
  3. 3.School of Psychology and Queensland Brain InstituteThe University of QueenslandBrisbaneAustralia

Personalised recommendations