Functional magnetic resonance imaging (fMRI) detects localized neuronal activity when a subject is presented with a task. Neuronal activity leads to local vasodilation, followed by an increase in blood flow, which increases the ratio of oxyhemoglobin to deoxyhemoglobin, resulting in increased blood oxygen level dependent (BOLD) signal intensity. Changes in BOLD signal between task and baseline states are used as a quantitative measure of activation to characterize and map sensory, motor, and cognitive function. Furthermore, differences between activation in patients and controls can be investigated using fMRI.

This book critiques, debates, and discusses some of the main issues associated with neuroimaging research, in particular the field of fMRI. The book originates from controversies and debate brought up at a workshop for neuroimaging experts and philosophers. It includes discussions on the characteristic nature of fMRI that makes this field of research challenging, and the methodological problems associated with fMRI analysis. It also discusses issues associated with locating brain areas for different functions, defining networks and interactions. The authors present their cases, engaging in debate over controversial issues, and it is left up to the reader to decide. This review will summarize the main points brought up in the book, and again, will leave the reader to form an opinion as to which analyses are best to address the particular research questions they have for their dataset.

In the first section “Location and representation,” several authors debate and discuss functional localization, and multivariate pattern analysis. The second section, “Inference and new data structures,” focuses on statistical issues including non-independence of samples, lack of cross validation procedures, multiple comparison correction, and hypothesis testing. The use of words, pictures, and graphs to present fMRI results are then discussed. The third section “Design and the signal” begins with an introduction to resting-state fMRI analyses, then discusses experimental design issues related to inference. High-resolution fMRI methods are covered, then the problems associated with inter-subject differences in anatomy and functional responses are discussed. The final section “Under-determination of theory by data,” takes a more philosophical position on neuroimaging. There is discussion about how pictures can sometimes misrepresent results. Attempts are made to answer questions relating to what we are trying to measure with fMRI, what neuroimaging actually represents, what conclusions can be drawn from neuroimaging research, and indeed why we use brain imaging in the first place.

fMRI studies involving functional localizers require the analysis to be split into two parts: First, whole brain analysis is used to identify regions activated by the task, referred to as functional regions of interest (fROIs); Second, analysis of the same data is restricted to the previously identified fROIs, to compare conditions and interactions. In the first chapter, A Critique of Functional Localizers, Friston and colleagues discuss the disadvantages of functional localizers, and suggest an alternative: to embed a functional localizer as an orthogonal contrast within the main experiment, using a factorial design. The authors aim to point out (mainly to reviewers who assert that functional localizers are the standard practice) that, although functional localizers have their place, there is often a better approach.

Saxe et al. provide a response to Friston et al.'s critique of functional localizers in Chapter 2, Divide and Conquer: A Defence of Functional Localizers. They discuss the advantages of using functionally defined ROIs, including the ability to find functionally equivalent brain regions across individuals so that data may be combined, and statistical power increased. An attempt is make to address the concerns of Friston and colleagues. While acknowledging the advantages of factorial designs, it is put forward that they should be used to complement independent localizer contrasts rather than replace them. The authors conclude that fROI analyses are useful to study phenomena for which functional regions have been well-described, can be easily identified, and are anatomically restricted.

In Chapter 3, Commentary on Divide and Conquer: A Defence of Functional Localizers, Friston and Henson give quite a convincing rebuttal to the assertions of Saxe et al. The authors point out that fROIs are limited in the questions they can address, and will most likely be undermined in the future by high-resolution fMRI and multivariate characterizations of fine-scale responses. They reiterate their stance on factorial designs and the lack of an advantage to using separate localizers. The key advantages of fROI analyses that Saxe et al. pointed out in the previous chapter are deconstructed. An important misconception about factorial designs is addressed from a statistical standpoint: the test for one main effect in a factorial design cannot bias the tests for other main effects or interactions just because the same data are used to find the ROI as to estimate the magnitude of the main effect.

Bunzl et al.'s Chapter 4, An Exchange About Localism, consists of a discussion between the three authors themselves, about localist versus holist strategies (or modularity vs. non-modularity). Lesion studies and the usefulness of fMRI in surgical planning are used as evidence against an entirely “holist” view. On the other hand, it is put forward that the brain is a network, and therefore functions can't necessarily be specifically localized.

Haxby introduces the reader to a multivariate approach to fMRI data analysis in Chapter 5, Multivariate Pattern Analysis of fMRI Data: High-dimensional Spaces for Neural and Cognitive Representations. Univariate analyses based on the general linear model can compare conditions only in a single dimension of activity strength, and are therefore limited to simple questions aimed at finding which functions “activate” a region. Rather, multivariate pattern analysis compares conditions in high dimensional space, characterizing activity by a vector. Multivariate pattern analysis works by model-based prediction, and is able to quantify the similarity of distinct patterns of neural response, in addition to detecting if the distribution among patterns is significant. Haxby argues that, since we assume that the brain can represent a variety of cognitive states and can vary locally in response strength, multivariate pattern analysis, therefore, has greater explanatory power.

The second section of the book begins with Chapter 6, Begging the Question: The Non-independence Error in fMRI Data Analysis. Non-independence error occurs when voxels are selected in which one condition produces greater signal change than the other, then the same data are used to evaluate the difference in signal change between the two conditions, thus the second analysis is not independent of the selection criteria. This error is common in fMRI because the data are so multidimensional. The authors, Vul and Kanwisher, give examples of non-independence error from the literature so that readers are better equipped to identify these errors when planning experiments, writing papers, and reviewing for journals. They also bring up the misleading practice of plotting (rather than testing) significant signal change in voxels that have been selected based on that same signal change. The authors provide suggestions to avoid non-independence errors: use of different datasets for selection and analysis; determining a priori whether selection criteria and analysis are independent; or assessing independence by numerical simulation. Regarding the Friston et al. versus Saxe et al. debate, Vul and Kanwisher agree with Saxe et al. that it is safer to use independent data. However, they also acknowledge that factorial designs are useful if they are truly orthogonal.

Poldrack and Mumford respond to Chapter 6, in a contribution entitled On the Proper Role of Non-independent ROI Analysis: A Commentary on Vul and Kanwisher. They begin by pointing out their common view with Vul and Kanwisher that the use of non-independent ROIs can lead more type I error when these analyses are used for inference, as well as inflated effect sizes in resulting plots. However, they believe there is still an important place for non-independent analyses in fMRI research. For instance, when used for quality control rather than inference, the reader can be assured that the whole brain effect is “real” and not driven by outliers. Also, plots can provide insight into the nature of interaction effects. Poldrack and Mumford refer readers to their article presenting guidelines for describing methods in fMRI including how ROIs are defined (Poldrack et al., 2008).

Vul and Kanwisher contribute Chapter 8, On the Advantages of Not Having to Rely on Multiple Comparison Corrections, which is a response to Poldrack and Mumford's chapter. Again, they begin by agreeing that non-independent display of data can be used for quality control but not inference, and that if whole brain analyses adequately estimate significant effects, is not necessary to collect additional independent data, thus solving the problem of multiple comparison correction. They outline and evaluate correction methods for multiple comparisons, including Bonferroni, false discovery rate and cluster size correction, but urge the reader to take care not to misapply these procedures. They still maintain that using an independently selected ROI approach will remove the additional complexity of multi-comparison correction, and is therefore worth extra data collection.

In their contribution in Chapter 9, Confirmation, Refutation and the Evidence of fMRI, Mole and Klein discuss how far neuroimaging can go in providing evidence to answer questions in cognitive science. The authors formulate general principles about what is required for neuroimaging data to prove cognitive hypotheses, and identify examples that fail to adhere to these principles. They point out that if data are “consistent” with a hypothesis, it does not necessarily prove the hypothesis is true, rather data need also to refute the contradictory null hypothesis. Again, examples of research that falls into this trap are given. The authors conclude that neuroimaging data are ill-suited to confirm cognitive hypotheses, can only provide evidence to support already well-substantiated theories, and therefore cannot be revolutionary.

Harman makes a brief note in Chapter 10, Words and Pictures in Reports of fMRI Research, to point out how readers of research reports can misinterpret conclusions based on misunderstandings about how certain words are used. They give the example of the phrases: “the data are consistent with the hypothesis” versus “the data confirm the hypothesis”. Harman also points out that readers may fail to appreciate that fMRI activation maps are not photographs, hence reading more into the results than they should.

Hanson and Glymour, in their contribution, Discovering How Brains Do Things, acknowledge that fMRI is the best available option to investigate the active brain, allowing brain function to be matched with brain structure and neural activity. They focus on some of the potential methodological problems and solutions for extracting information about how brain regions influence each other from fMRI, that is, effective connectivity. The authors discuss the steps involved in constructing fMRI variables, including identifying sets of voxels, clustering, and spatial smoothing. There is also discussion on how a brain intervention (stimulus) can directly or indirectly change a subject's pattern of brain activation, as indicated by the BOLD response in the fMRI signal. Hanson and Glymour consider what it means to claim that one region of activity is influencing another. Finally, the authors discuss methods for extracting information about interactions between brain regions during a cognitive task, and how these methods can be tested and evaluated.

The third section of the book begins with Biswall's Chapter 12 on Resting-state Brain Connectivity. Resting state connectivity is referred to as spontaneous, low frequency signals that fluctuate synchronously between functionally related regions, in the absence of any stimulus or task. Some suggest that it represents a default mode of information processing. Biswall discusses the origin of resting state connectivity, presenting the biophysical origin hypothesis, and the cognitive origin hypothesis. He examines studies where resting state connectivity and task induced activity are compared, concluding that task-induced activation maps may underestimate the size and number of functionally connected areas, while resting state connectivity analysis may more fully reveal functional networks. The author believes there's a high likelihood that analysis of resting state physiological fluctuations will contribute to clinical practice.

Chapter 13, Subtraction and Beyond: The Logic of Experimental Designs for Neuroimaging, is another contribution by Poldrack. This chapter characterizes the experimental designs used for fMRI, focusing on how a design can limit the conclusions that can be made. Poldrack defines the most basic and commonly used experimental design, the subtraction method, and critiques its underlying assumptions and limitations. Alternatives to the subtraction method are presented including cognitive conjunction analysis, factorial designs, parametric designs, and priming/adaptation designs, which are also associated with various limitations and assumptions. The author considers the advantages of event related designs over blocked designs, but also discusses the cautions to be taken. Generally, it appears that experimental designs in fMRI need to be formulated based on an understanding of how the task relates to the underlying cognitive processes.

Grill-Spector's Chapter 14, Advancements in fMRI Methods: What Can They Inform About the Functional Organization of the Human Ventral Stream?, reviews methodological advances in fMRI that have been used to clarify the neural representations underlying object recognition, including fMRI adaptation, multi-voxel pattern analysis, and high-resolution fMRI. There are limitations to using standard-resolution fMRI to understand neural representations of faces and objects, as fine-resolution neural organization cannot be resolved. The author highlights how these advanced methods have been used to extend theories of object recognition, the most studied region being the fusiform face area, and believes that a combination of these methods will be most profitable for understanding neural representation of objects and faces.

Chapter 15, Intersubject Variability in fMRI Data: Causes, Consequences and Related Analysis Strategies, by Poline et al. addresses the challenge of how to summarize, describe and interpret results from several subjects in group comparisons of fMRI data. Despite the difficulties of inter-subject variability, often it is worth persevering in order to obtain the rich source of information it contains for understanding brain function. This chapter points out the limitations in current analysis approaches, and suggests techniques and strategies to help address inter-subject variability issues, including: assessing group homogeneity with various diagnostic tools and techniques; parcellation techniques to account for small spatial mis-registration; and analyses based on reproducibility measures and large databases.

The final section begins with Roskies’ Chapter 16 entitled Neuroimaging and Inferential Distance: The Perils of Pictures. Roskies writes about the inferential steps between observations and the conclusions drawn from them, referred to as “inferential distance.” The steps from neural activity to raw data to MR image are referred to as the causal stream, while the link between neuroimaging data, cognition, and behavior is referred to as the functional stream. The author discusses the danger of biased conclusions and mis-interpretation when inappropriately presenting fMRI images as photographs of brain activity, rather than representations of the location and level of neural activity. Roskies asserts that the inferential distance between neural activity and the fMRI activation map may generate misconceptions, particularly in the public arena.

Loosemore and Harley provide a rather negative view on neuroimaging in Chapter 17, Brains and Minds: On the Usefulness of Localization Data to Cognitive Psychology. They do not believe that brain imaging (specifically in relation to localization of brain function) can teach us anything interesting about cognitive psychology, and do not seem at all impressed with the “obsession” with brain imaging which is “taking over the world.” They view neuroimaging as a fun but expensive exercise where the images may look good, but do not add anything to science. They base their argument around some rather non-mainstream studies (some of which are not neuroimaging studies), which they chose in order to illustrate how neuroimaging in general cannot give any new information about the structure or function of any mechanisms in the human cognitive system. They evaluate the chosen studies against a theoretical molecular model of cognition. They conclude that, through neuroimaging, “we are being flooded with accurate answers to questions about the brain location of mechanisms that we do not believe in and inaccurate answers to questions about the brain location of mechanisms that are currently not terribly interesting. This state of affairs seems to us to be a great leap backwards” (p. 240).

Bechtel and Richardson take the middle ground on neuroimaging in Chapter 18, entitled Neuroimaging as a Tool for Functionally Decomposing Cognitive Processes. The authors explain the contribution of neuroimaging to science, which when done well, can help to understand and explain mental phenomena. They address the key criticisms that Loosemore and Harley made in the previous chapter, and provide a contrasting conception of how science progresses. Although this chapter focuses on rebutting the critics of fMRI, the authors also recognize that fMRI studies are often over-interpreted by enthusiasts. In a balanced view, they state that fMRI is just one of many technologies that will become more useful as fMRI image quality and resolution increases and analysis techniques improve.

The final chapter by Coltheart, What Is Functional Neuroimaging For?, sums up the contributions of the book in terms of the goals of fMRI: neuroanatomical localization of cognitive processes; testing theories of cognition; and, testing neural models. The author proposes four helpful questions to ask of any existing or proposed fMRI study: If the goal is functional localisation, ask “what well-accepted model of cognitive processing in that cognitive domain does the study presume?”; If the goal is to test a model of cognitive processing, ask “what plausible outcome of the neuroimaging study might have been obtained that would have counted as evidence AGAINST that model?”; If the goal is to adjudicate between competing models of cognitive processing, ask “what is it about each model that is contradicted by the other(s) – are the models genuinely incompatible?”, and for each competing model ask “what plausible outcome of the neuroimaging study might have been obtained that would have counted as evidence against that model?”. Like several other contributors to the book, their view is that fMRI has provided new and interesting information about the brain, but not about cognition.

For neuropsychologists and other researchers who are considering including an fMRI component into their research projects, Foundational issues in Human Brain Mapping would be beneficial. It provides a comprehensive guide to the methodological issues associated with fMRI, and the interpretation of results obtained from fMRI experiments. However, some of the discussion is quite technical and may not be easily understood by an audience that lacks in-depth understanding of fMRI.

I gratefully acknowledge Dr. Christopher L Adamson for providing feedback on this review.