(1)
Department of Neurosurgery, St Elisabeth-Tweesteden Hospital, Tilburg, The Netherlands
Functional neuroimaging techniques, first positron emission tomography (PET) and later functional MRI (fMRI), have revolutionized cognitive neuroscience. These tools have also greatly improved our understanding of how language is implemented in the brain. Almost from the beginning, fMRI was also applied for language mapping in surgical practice because of its obvious benefits: high-resolution whole-brain mapping without the need for invasive procedures. Other clinical applications that have been investigated, although less frequently, are the use of fMRI as a tool to help diagnose or understand diseases that lack clear neuroanatomical characteristics or as a predictor for language outcome after stroke (see Chap. 9).
But in contrast to its success in the cognitive neurosciences, it has proven to be more problematic than expected to turn fMRI into an instrument that clinicians want to use for their patients. This chapter explores fMRI from both a neuroscientific and a clinical perspective. First, key principles of functional neuroimaging are listed, together with some methodological considerations. Next, the history of functional neuroimaging is described and the scientific progress that has been made in understanding language with the help of neuroimaging. Finally, we will explore the potential of fMRI in a clinical setting and discuss strategies to further develop it as a clinical instrument.
8.1 Brief Introduction to the Method
Before the 1980s, the neural basis of brain functions was largely inferred from the results of post-mortem lesion studies and the small number of neurosurgical patients that had been investigated with electrocortical stimulation mapping. The problem was that this data had been gathered from diseased and damaged brains, and the presence of the lesion had probably already altered the functional organization of the brains of these patients or at least induced some form of compensation. PET provided, for the first time, an instrument that could study the normal functional neuroanatomy of brains unaffected by pathology or potential functional reorganization with good spatial precision [1]. PET is a slightly invasive technique because it requires the intravenous injection of radioactively labelled water (with a half-life of 2 min). At the beginning of the 1990s, it was discovered that changes in brain activity can be visualized without any contrast agent, using an MRI machine. This technique became known as ‘functional’ MRI or fMRI. The first fMRI studies were published in 1992 by Bandettini [2], Ogawa [3] and Kwong [4], and these groups are usually credited with the discovery.
The principle upon which PET and fMRI rest is that neurons that become more active increase their energy consumption. The brain is a particularly greedy consumer of energy, as it takes up to 20% of the bodily energy consumption with only 2% of its body weight [5]. To supply this energy, local blood vessels dilate to increase the inflow of nutrients. This leads to regional changes in the brain’s physiology and metabolism, of which some can be mapped with surprisingly precise spatial resolution. Current functional imaging techniques mostly use differences in the concentration of glucose, oxygen or haemoglobin as an indirect means to construct three-dimensional images of brain activity.
fMRI is the only neuroimaging technique that is completely non-invasive. It exploits the fact that the actual inflow of oxygen to regions with increased neuronal activity is much more than locally used by neurons [6, 7]. Malonek and Grinvald very eloquently described this as the brain ‘watering the entire garden for the sake of one thirsty flower’ [6]. This phenomenon is also called the ‘blood oxygen-level dependent’, or ‘BOLD’ effect (Fig. 8.1), and is the basis of the contrast that is used in almost all fMRI measurements to visualize changes in brain activity.1
Fig. 8.1
The exact relationship between neural and cerebrovascular changes remains unknown, but microelectrode recordings in both animals and humans strongly suggest that the BOLD signal corresponds to local field potentials (LFPs) [8, 9]. LFPs reflect the input and intracortical processing of a population of neurons, rather than their output. Several studies have also reported a correlation between fMRI signals and an increase in the power spectrum as measured with electrocorticography for a variety of tasks [10, 11]. Note that a typical voxel (3–5 mm in each dimension) contains over 5 million neurons, 50 billion synapses and 200 km of axon [12]. (a) MRI scan shows the position of a microelectrode within the visual cortex of a monkey (Logothetis 2001) [9]. (b) Results of an fMRI experiment in the same monkey with visual stimuli (rotating chessboard for 10 s). BOLD responses (colour) are measured from the visual cortex. (c) Graph shows that the BOLD response starts several seconds after neural activity within the same region. Logothetis (2001) concluded from the experiments that ‘results show unequivocally that a spatially localized increase in the BOLD contrast directly and monotonically reflects an increase in neural activity. (…) In a first approximation BOLD and neural responses are shown to have a linear relationship for short stimulus presentation durations’ [9] (Figures from Logothetis (2001) [9])
8.1.1 Task Conditions
fMRI provides the investigator with a three-dimensional map of signal changes in the brain. These changes can be spontaneous or associated with different task conditions (see Fig. 8.2 for an example of a block-designed experiment). Language studies have often focused on the distinction between productive and receptive functions, but this dichotomy has not always been helpful in designing fMRI experiments to unravel the anatomical basis of language functions, nor in understanding aphasic problems. Binder suggested the use of a classification from linguistics for fMRI experiments, which splits up language into five subcomponents instead of two:
Fig. 8.2
BOLD-fMRI language experiment in a patient with a left temporal lobe tumour (low-grade glioma located predominantly in the left superior temporal gyrus). Maximal surgical removal of such a glioma requires information on critical language-related structures located nearby. BOLD-fMRI can only measure relative signal changes and does not provide the investigator with absolute measurements of brain activity. As such it relies upon a contrast of two or more experimental conditions (i.e. it requires a reference condition against which signal changes can be measured). In its simplest and most commonly used form, as shown here, the experiment has a ‘block-design’ with two conditions. In the ‘active’ condition, single nouns were presented, and the patient was instructed to think of a related verb (e.g., ball → throw, car → drive). In the control condition, symbols were presented; it was hypothesized that these stimuli do not engage the language system. There were ten epochs of 30 s (five for the active and five for the control condition). Spatial resolution (voxel dimension) is 3 × 3 × 3 mm. Red voxels denote areas where the signal is significantly higher during the active than during the control condition. Note that the results from all different stimuli condense in a single functional brain map. To provide an anatomical frame of reference, fMRI results are projected onto detailed anatomical MRI images (here, a cortical rendering of a T1-weighted MRI scan). In this patient, considerable activation of frontal and temporal language areas in both hemispheres was observed, suggesting bilateral representation of language functions. (Sub)cortical mapping during awake surgery confirmed language areas in the left hemisphere (for obvious reasons, no information was acquired from right fronto-temporal areas during surgery). These single-task fMRI experiments are increasingly used for the planning of brain surgery. Point of discussion is to what extent fMRI areas are truly critical for normal language functions and have clinical relevance. In other words, it is not known whether removal of these areas will lead to lasting language problems for patients
(1) phonetics, the processes governing production and perception of speech sounds; (2) phonology, the processes by which speech sounds are represented and manipulated in abstract form; (3) orthography, the processes by which written characters are represented and manipulated in abstract form; (4) semantics, the processing of word meanings, names, and other declarative knowledge about the world; and (5) syntax, the processes by which words are combined to make sentences and sentences analysed to reveal underlying relationships between words [14].
Although this classification seems a better approach for the study of language, it should be noted that it cannot always be expected that different brain functions add up linearly. Cognitive processes can interact in a complex and non-linear fashion, creating unpredictable effects [15]. Block-designed experiments are rooted in a form of modular thinking and rely on the principle of pure insertion, implicitly assuming that perception, cognition and action are largely independent brain processes [16]. Modern theories, on the other hand, propose that the same neurons can be active during both perception and action and that sensorimotor systems form an integral part of cognitive processes [17].
8.1.2 Detection Power
The signals that can be picked up with fMRI only increase a few percent after brain regions become neurally more active. These changes are so small that they easily get lost in huge amounts of background activity [13, 18]. There is an ongoing discussion about the exact physiological processes that take place when the brain is engaged in certain tasks [19]. The term active or activated is frequently used in functional neuroimaging studies to denote an area that is responsive during a particular sensorimotor or cognitive task. The term implies that brain areas change from a dormant state into a functional one. However, brain areas are probably never truly ‘inactive’. In fact, most of the brain’s energy is needed to sustain a certain baseline neural activity. It is therefore better said that areas become more (or less) active and that the level of activation—whatever that means—varies over time.
Investigators may use different strategies to increase the detection power for the task-related signals that they are interested in. These come at a cost, though, and impose various constraints on the design of experiments. Common strategies include repetition of stimuli and spatial smoothing of the data.2 What raises the complexity of fMRI analyses further is that signals are measured from several thousands of voxels and that for each voxel hundreds of measurements are performed in a time series. As is the case with noisy data, statistics are subsequently needed to decide whether or not the measured signals contain information. Functional brain maps are therefore, by definition, statistical representations of the outcome of the experiment.3 Although there are a number of popular software packages for fMRI analysis (e.g. SPM, AFNI, FSL), standardization is lacking at this point [22].
fMRI can be performed with a regular MRI scanner, which is nowadays widely available in hospitals. At most, it requires relatively inexpensive adaptations to software and hardware, making it much more accessible than PET. Current state-of-the-art fMRI techniques are sensitive enough to pick up signals related to a single stimulus in individual subjects at a millimetre resolution. However, a typical cognitive fMRI experiment provides brain maps reflecting average activity of groups of subjects with a resolution of about 1–2 cm.
In a relative short time, fMRI has become an immensely popular tool in neuroscience. fMRI protocols and techniques have continuously evolved, and by 2010, scientists published about it in more than 1500 articles [24]. Other brain mapping techniques, such as electroencephalography (EEG) or magnetoencephalography (MEG) are used much less frequently because of their lack of spatial precision (EEG, MEG) or very high costs (MEG). In stark contrast, clinical fMRI investigations have not changed their basic methodology much since the pioneering language studies of Petersen or Binder or the first neurosurgical patient studies that came out in the 1990s [25, 26]. This is remarkable because there have always been significant discrepancies between the results of fMRI and those of the invasive clinical techniques. One would have expected this to be a strong impulse to re-evaluate both new and established clinical methods and to understand why these results are so different (as they both, of course, aim for a similar goal: a map of critical language areas). In reality, we think, relatively little effort has been made to close this barrier. Clinicians still rely on the Wada test and direct electrical stimulation (DES) for language mapping, although they are starting to acknowledge that these techniques have significant flaws and that they are not the ideal gold standards against which to judge fMRI maps (see also Chap. 6) [1, 14]. Neuroscientists, on the other hand, never really turned their attention to the mapping of functions in individuals—an important prerequisite for clinical use. For some reason, the same fMRI methods that were developed for group studies were also used for precise localization of functions in individual patients. However, single-subject fMRI requires a different approach. Before this discussion is entered in more detail, we will first summarize what neuroscientific studies, and in particular those that used functional imaging techniques, have taught us about the neural basis of language.
8.2 Historical Perspective
The underlying assumption of BOLD imaging is that regional blood flow changes are correlated to local neural activity. Such a relationship was already suspected at the end of the nineteenth century [27, 28]. Angelo Mosso (1846–1910), an Italian physiologist, was the first to actually gain experimental evidence of this phenomenon [29]. He measured pulsations of the brain in a patient with a skull defect after a neurosurgical intervention. A farmer named Bertino had bone pieces removed after a traumatic skull fracture, and that had left him with a bony opening in the skull. Mosso designed a device that could simultaneously measure pulsations from the forearm and from the skin overlying the bone defect. He hypothesized that these latter measurements reflected changes in brain volume and intracranial pressure. Then, and this was truly a scientific inquiry, Mosso asked Bertino to perform certain cognitive tasks (i.e. multiplication) and noted that—after some delay—the pulsations from the surface of the brain increased, whereas those of the forearm did not (Fig. 8.3). This indicated, or at least suggested to him, that blood rushes to the brain when mental performance increases. As such, he was one of the first researchers to develop a technique for in vivo brain imaging. Mosso also performed other experiments that—as we read it now—require somewhat more wishful thinking:
Fig. 8.3
Mosso’s device for simultaneously recording pulsation in the arm and brain. The brain showed stronger pulsations after an event that stimulated brain activity (shown with arrow) [29]
The subject to be observed lay on a delicately balanced table which could tip downwards either at the head or the foot if the weight of either end were increased. The moment emotional or intellectual activity began in the subject, down went the balance at the head-end, in consequence of the redistribution of blood in his system.4
At the time, many other researchers were interested in the psychophysiological relationships between the mind and brain. Some, including Mosso, studied temperature changes of the brain and performed various experiments in animals and man to study the causal effects of these changes. Mosso described them in his book Die Temperatur des Gehirns (1894). He concluded that ‘the fluctuation in the temperature of the brain was independent of blood temperature and was likely related to the metabolic activity of the brain itself’ [29]. There are several historical notes of interest here. For instance, Hans Berger (1873–1941), the inventor of electroencephalography, was inspired by Mosso’s work. In his clinic in Jena, he adapted Mosso’s technique of plethysmography and this eventually led to development of the EEG. As Schiller wrote: ‘One day in 1924 this none too rewarding pursuit [of Berger] gave him the idea of using electrodes, to replace “thermoencephalography” with the E.E.G.’ [30]. Broca also tried to localize brain lesions via recording of the temperature of the skin. He had learned from his general surgical work that blood flow changes affected the temperature of a limb. Broca used this information to determine at what exact place a diseased extremity should be amputated. In an analogous way, he proposed that brain lesions led to changes in the cerebral vasculature and subsequently to changes in local temperature. To measure these changes, Broca had devised a ‘thermometric crown’ that he considered sensitive enough to pick up temperature changes despite the physical boundaries of the dura, skull and skin. Broca also hypothesized that brain temperature should increase with the execution of cognitive tasks, in particular in the frontal areas, and he described some of these experiments in his work [31]. Again, these were among the first attempts at functional brain imaging.
Following up on these pioneering studies at the end of the nineteenth century, Charles Roy (1854–1897) and Charles Sherrington (1857–1952) studied brain volume changes in animals via a more sophisticated measuring device that was implanted in the skull of an animal (Fig. 8.4). This allowed them to record these changes under various controlled experimental circumstances, for instance, during the stimulation of peripheral nerves or medulla oblongata, the restricted inflow or outflow of blood to the cranium or asphyxia. This resulted in a landmark paper that was published in 1890 and that described some of the basic principles of cerebral blood flow regulatory mechanisms [33]. Their studies supported a model whereby cerebral blood flow is controlled by both ‘extrinsic’ factors (arterial and venous blood pressure) as well as ‘intrinsic’ (local) factors:
Fig. 8.4
‘A trepan hole, about 22 mm in diameter in the case of dogs, but smaller when a cat or a rabbit was used, was made as near the middle line of the vertex of the cranium as is compatible with avoidance of the longitudinal sinus, after which the subjacent dura was removed by a circular incision. After any oozing of blood from the diploë had ceased, a small metal capsule, of a size corresponding with that of the trepan hole, was fixed over the aperture by means of screws. The shape of the capsule and the mode of fixing it firmly to the skull can be seen on reference to Fig. 1. The lower opening of the bell-shaped capsule (a) is closed by a very flexible, delicate, animal membrane (e), of the kind already used by one of us (R.) in other apparatus. It is tied on in such a way that it readily follows all changes in the level of the part of the cortex on which it rests, while it prevents any escape of the air with which the capsule is filled. Outside the capsule, about two mm from its lower edge, is a projecting rim (b), which rests on the external surface of the cranial bone. This rim has in it two notches, in which fit two metal pins (c and d), bent at right angles at their lower ends, so that they can hook under the bone on opposite sides of the hole. By means of small thumbscrews on the upper parts of these pins, the capsule is held firmly in position. The upper opening of the capsule is connected by means of rigid-walled tubing with the recording apparatus. This latter consists of an arrangement similar to that which one of us has described as useful for studying the form of the pulse wave and which is shown in Fig. 1a. A light piston, escape of fluid by the side of which is prevented by a flexible membrane of the kind already referred to, conveys to a recording lever any changes in the volume of the brain (Text and figures taken from Roy and Sherrington (1890)’ [33])
These facts seem to us to indicate the existence of an automatic mechanism by which the blood supply of any part of the cerebral tissue is varied in accordance with the activity of the chemical changes which underlie the functional action of that part. Bearing in mind that strong evidence exists of localization of function in the brain, we are of opinion that an automatic mechanism, of the kind just referred to, is well fitted to provide for a local variation of the blood-supply in accordance with local variations of the functional activity. (Roy & Sherrington, 1890 [33])
The principles that were laid out by Roy and Sherrington are still the basis of modern functional neuroimaging techniques, although controversy remains about the exact underlying mechanisms. See for an overview Fox (2012) [7].
8.3 What Neuroscientific Studies Taught Us About the Neural Basis of Language
An important goal of neuroscience is to unravel and, if possible, to understand the neural architecture that underlies brain functions. The traditional approach, that started with the lesion-deficit studies at the end of the nineteenth century, is to search for consistent and meaningful structure–function relations among subjects. This is also the approach that is still often used in functional imaging studies. Data from individual subjects are transposed to some standard brain template, after which they are averaged to form groups. Averaging is beneficial in the sense that it reduces the influence of noise and individual variations that are considered to be of no interest. Although meaningful differences between subjects may get lost in this process, this trade-off is generally accepted by the neuroscientific community in search for overarching theories and models [34].5
Functional neuroimaging studies have greatly improved our understanding of how language is implemented in the brain, and have provided alternatives for the classic convictions on language localization. It is worth taking a closer look at some of the older studies first, as these already yielded several observations that conflicted with the classic Broca-Wernicke model, both from a conceptual and anatomical point of view.
8.3.1 Some Landmark Studies
One of the first functional imaging studies on language processing was published in 1988 by Petersen and colleagues: Positron emission tomographic studies of the cortical anatomy of single-word processing [37]. In the introduction of their paper in Nature they refer to Geschwind’s model as the clinical model that was most widely accepted at that time, and that ‘argues for serial processing, with an early recoding of visual input into an auditory-based code which is used in turn for semantic and articulatory access’ [37]. The results from Petersen’s functional imaging studies did not support this model, but were more consistent with models of parallel processing that had already been suggested in contemporary studies [38, 39]. The authors concluded that there were three main findings that were inconsistent with any model of serial language processing:
First, there is no activation in any of our visual tasks near Wernicke’s area or the angular gyrus in posterior temporal cortex. Visual information from occipital cortex appears to have access to output coding without undergoing phonological recoding in posterior temporal cortex. Second, tasks calling for semantic processing of single words activate frontal, rather than posterior, temporal regions. Third, sensory-specific information appears to have independent access to semantic codes and output codes; simple repetition (output tasks) of a presented word failed to activate the left-frontal semantic areas (association tasks) [37].
Petersen studied both auditory and visual processing of single words. They used four behavioural conditions in a three-level hierarchical block design. ‘Each task state was intended to add a small number of operations to those of its subordinate (control) state’ [34]. A description of the tasks and results is given in Fig. 8.5. The authors proposed a model whereby there are multiple routes between areas that code articularly, phonological or semantic information. The importance of the early PET studies was far-reaching, as, for instance, stated by Price in an extensive review of 20 years of PET and fMRI language studies (1992–2011) [40].
Fig. 8.5
Schematic results of a PET language experiment in 17 healthy subjects from the paper of Petersen (1988) [37]. There are four behavioural conditions, with both auditory and visual presentation of stimuli (looking at fixation point—listening or reading passive words—repeating visually or auditory presented words—generating a verb from a given noun). The researchers calculated three different contrasts from these task conditions (the right column in the table shows the cognitive processes that are hypothesized to be different in the stimulated state versus the control state). The figures denote the ‘activated’ areas that were found
They [the first PET studies] illustrated that functional imaging could provide anatomical localization with a precision that far exceeds that attainable with human brain lesion studies. Moreover, the study of healthy subjects avoids possible confounding effects of brain lesions, such as compensatory reorganization of brain function [41–43]. Methodological challenges were also well appreciated, particularly when the results appeared to contradict classic axioms of language organization. For example, Steinmetz and Seitz (1991) [44] argued that data should not be averaged over subjects because intraoperative stimulation showed diversity in location of language functions and morphometrical imaging studies showed diversity of brain shape and gyral patterns that would be difficult to correct with anatomical normalization techniques. Many other concerns were succinctly addressed in a review by Petersen and Fiez (1993) [45], who pointed out that functional neuroimaging results should be viewed as evolutionary, rather than revolutionary and that they were most interpretable when they were backed up by supporting data from other studies.(…) Petersen and Fiez (1993) [45] also emphasized that complex language functions were not localized in specific brain regions; they were distributed across networks of regions with each area making a specific contribution to the performance of the task, which depends on its connections to other areas in a parallel distributed hierarchy. In this context, understanding the functional anatomy of language cannot be deduced from a single experiment; rather, it requires the integration of results from multiple experiments using multiple techniques [40].
When fMRI was further developed, it gradually became the most frequently used tool to study the neural basis of language and other cognitive functions. One of the first fMRI studies that targeted language areas was from Binder and colleagues (1997) [46]. His group studied 30 right-handed volunteers with a semantic decision task, with auditive presentation of words (via MRI-compatible headphones). There were two control conditions: one in which subjects had to perform a tone decision task and one where they were asked to remain relaxed and motionless (i.e. without explicit instructions) (see Fig. 8.6 for details and results). As in Petersen’s and other studies, some of the results were clearly incongruent with the classic language view [37, 47, 48]. For instance, despite abundant activation of left temporoparietal areas on the group maps, Wernicke’s area was not clearly activated. Most of the temporal activation was found in the middle temporal gyrus. Another remarkable finding, also found in Petersen’s study, was that the semantic decision task not only activated temporoparietal areas, as expected from classic teaching, but also left frontal language areas. In general, frontal areas seem more easily activated during fMRI language tasks than temporoparietal areas. These frontal areas often extend well beyond the classic Broca’s area and include large parts of the medial and lateral prefrontal cortex.
Fig. 8.6
Block-designed fMRI language experiment taken from the paper of Binder (1997) [46]. Experimental conditions included a ‘rest’ state and two behavioural tasks. Stimuli were given via headphones and were either tones or sampled male speech sounds. In the semantic decision condition, subjects had to decide (via a button press) whether a spoken English noun was an animal that was both ‘native to the United States’ and ‘used by humans’. In the control condition, two different tones were presented (500 and 750 Hz). Subjects had to respond when they heard two consecutive 750 Hz tones. (Top) Group results for 30 right-handed healthy volunteers. fMRI activation maps are shown for the semantic decision versus tone decision comparison, whereby the results were scaled to an averaged standard brain. Note strict left-sided lateralization and extensive involvement of large parts of frontal, temporal and parietal areas outside classic language areas. Yellow-red colour scale denotes the probability that voxels are activated in the semantic decision task relative to the tone decision task. Cyan-blue voxel scale with the reverse contrast (note that these areas are strongly lateralized to the non-dominant hemisphere). (Bottom) fMRI language areas in an individual subject (26-year-old male)
But judgements on how well functional neuroimaging results correspond to those of classic language models depend on the investigator’s perspective. In 2000, Price reviewed the functional neuroimaging studies that had been performed thus far—focusing on single-word processing tasks—and specifically compared them against classic nineteenth- and twentieth-century lesion-deficit models [49]. Although Price suggested modifications to the classic models, based on the findings from modern imaging studies and cognitive psychology, she concluded that there were more commonalities than differences:
The correspondence to the 19th Century neurological model illustrated in Figure 1 [our Fig. 8.7] is clear although a few refinements have been made. First, the site that corresponds to the function of Wernicke’s area is the upper bank of the posterior superior temporal sulcus. Second, the critical site for articulatory planning is the anterior insula, not the third frontal convolution (Broca’s area). Third, the angular gyrus is not specific to visual word forms but is engaged when semantic associations are made. Fourth, the meaning of words is also distributed along the left inferior and middle temporal cortices. Fifth, reading and name retrieval tasks activate the left posterior inferior temporal lobe. This region is thought to have monosynaptic connections to Broca’s area (DiVirgilio & Clarke, 1997) thereby providing the semantic reading route that was missing from the 19th Century model. In brief, the only anatomical regions that were missing from the 19th Century neurological model were in the inferior temporal cortices, areas that are relatively resistant to the ischemic damage that the lesion deficit model is dependent upon [49].
Price, as a neuroscientist, speaks of ‘only a few refinements’ [49]. Overall, functional imaging results indeed show overlap with those of the lesion-deficit models at a generic level (Fig. 8.7). However, when judged from a more clinical perspective, there are important differences. It is easily seen that some of the areas of Price’s model do not accord with clinical experience. For example, surgery within the left inferior temporal lobe generally does not result in language deficits, and electrical stimulation finds language functions in a far wider temporal region than the superior temporal sulcus alone. Neurologists are interested in language representation in the individual patient and differ in their questions from neuroscientists, who want to generalize results across populations. Neurosurgeons are even more exacting and require individualized information on ‘eloquent’ and ‘non-eloquent’ areas with sub-centimetre accuracy for the planning of their operations. They are interested in the precise anatomical organization of language and want to know what areas are truly essential for normal language functioning. In this respect, the generalized models that are drawn from functional imaging results are not very helpful.
Fig. 8.7
Proposed language model as derived from functional imaging studies by Price (2000) [49]. The model was considered to be largely consistent with the classic models of language processing. Semantics is described at word level
An important point—and warning—here is that most people will probably look at these fMRI images, or any brain map for that matter, with the implicit assumption that all of the highlighted areas play a critical role in normal language functioning. This is a very understandable but wrong assumption. These experiments were not designed, and thus not meant, to provide us with answers to such questions (nor may we assume that the uncoloured areas are not involved in language processing). Even for me, with a fair amount of background in functional neuroimaging, it is sometimes difficult to suppress this most intuitive reaction. We may not compare fMRI maps directly to those of lesion-deficit studies. Graphical images, as we have repeatedly seen before, easily speak for themselves (think of the schemes of Wernicke or Penfield). But it is in fact impossible to ‘see’ what they actually represent without sufficient background information.
8.3.2 From Single Words to Sentences
In the meantime, a wealth of fMRI studies has been published. A large part of these studies have investigated the processing of single stimuli, using tasks such as word generation or picture naming (see for some of the earlier reviews, for instance, Bookheimer (2002) [50] or Demonet (2005) [51]). Clinical studies predominantly stuck to these ‘simpler’ tasks, whereas neuroscientists and linguists moved further and also began studying the neuroanatomical basis of more complex tasks, such as sentence processing. For obvious reasons, the neural processes that are responsible for the production and understanding of sentences are even more complex than those of single words. In order to comprehend a sentence or an utterance, it is not enough to deduce the meaning of individual words in the linear manner to which they are presented to the listener or reader. The relationship between words, in terms of the overall meaning of a sentence, is often non-linear, as parts of sentences can be embedded or otherwise have a different hierarchical structure (e.g. the cat that chases the dog is black). Information therefore needs to be kept ‘online’ before its meaning can be grasped, and additional resources (such as selective attention and working memory) may be required to understand and apply grammatical and syntactical relationships at sentence level.
In 2006, Vigneau and colleagues published a meta-analysis of fMRI and PET studies that went beyond single-word processing [52]. They specifically investigated phonology, semantics and sentence processing. Despite the methodological limitations that are inherent to meta-analyses (the authors had no access to the raw data and had to deal with differences in spatial normalization and data analyses methods across studies), the authors claimed that spatial resolution was ‘under the gyral level’. Several functionally specialized networks were identified that covered extensive areas in the frontal and temporal lobes (see Fig. 8.8). A number of interesting propositions were extracted from these results. For instance, it was found that phonological and semantic processing have separate networks within the left inferior frontal gyrus, confirming earlier work by Poldrack (1999), who described an anterior–posterior dissociation of phonological and semantic areas [53]. Also, the semantic network was found to be much larger than traditionally envisioned, including the angular gyrus, superior and middle temporal gyrus, fusiform gyrus, temporal pole and clusters in the left inferior frontal gyrus. ‘This semantic network can be considered to construct an overall meaning on the basis of the association of integrated knowledge issued from the main domain of external (audition, vision) and internal (long-term memory, emotion) messages; this construction of sense forms the foundation of language communication’ [52]. Finally, the authors describe three different working memory loops: for phonology, semantics and sentence processing.
Fig. 8.8
Meta-analysis of 129 fMRI studies of healthy volunteers that specifically investigated phonology (blue), semantics (red) or sentence processing (green), as studied by Vigneau (2006) [52]. (Top) 730 activation peaks are shown on a cortical rendering of the left hemisphere (in stereotactic space). (Middle) For further analyses, peaks were clustered. The semantic network is shown, which includes a dorsal and a ventral component in the temporal lobe. The ventral component is dedicated to visual material and includes T3p at the interface between phonological and semantic processes for audio-visual processing (yellow). The dorsal component is dedicated to auditory material and includes the voice area (yellow) at the interface between phonological and semantic processing. In the frontal lobe, the semantic areas are located in the anterior part of the inferior frontal gyrus. (Bottom) Three different working memory loops. The working memory loop for phonological material is shown in blue and connects inferior frontal areas to those in the parietal lobe. The working memory loop for semantics (red) includes a frontal area at the junction of the precentral gyrus and opercular part of the inferior frontal gyrus (PrF3op) and the angular gyri. The working memory network for sentence and text comprehension includes the posterior part of the middle frontal gyrus (F2p) and the posterior part of the superior temporal sulcus (STSp, green)
8.3.3 A New Anatomical–Functional Perspective
The four studies that were briefly reviewed before (Petersen, Binder, Price, Vigneau) are exemplary for what is generally found in functional imaging experiments when these are carefully analysed and interpreted. These results are not sufficiently explained by the older lesion-deficit models and contrast with what is generally taught in medical school. In fact, they make a strong case that the neural basis of language needs to be redefined. One of the more consistent findings in both fMRI and modern lesion studies is that language production and comprehension are not restricted to, respectively, left inferior frontal and temporoparietal regions. Broca’s area is clearly involved in both production and comprehension, as was already observed in the early functional imaging studies of Petersen and Binder. A similar conclusion holds for Wernicke’s area; when defined as the posterior part of the superior temporal gyrus, this area seems predominantly involved in phonological processes that facilitate both language production and comprehension. Before words can be spoken and before relevant muscles are innervated for this purpose, the neural representations of speech sounds (phonemes) need to be made available. Phonological retrieval as well as temporary storage of phonetic sequences is thus an important prerequisite for normal execution of speech [54]. Phonological processing is also central to the ‘acquisition of long-term lexical memories of novel words’ [54, 55]. As such it plays a vital but indirect role in language comprehension. The actual meaning of words and sentences is represented in a much wider (and bilateral) network that is located outside of classic Wernicke’s area (Fig. 8.9). This view is perfectly in line with the original ideas of Wernicke and Lichtheim, but not very commonly held in today’s clinical practice, where lesions in Wernicke’s area are still largely synonymous with comprehension disorders [54, 56–58].
Fig. 8.9
Binder’s model (2015) of the major posterior language systems, based on modern lesion and functional imaging studies [54]. Classic Wernicke’s area (posterior temporal gyrus) is not directly associated with verbal comprehension, but instead with phonological processing. ‘Yellow indicates a bilateral speech phoneme perception system. Blue indicates the Wernicke area, which supports prearticulatory phonologic retrieval. Red indicates the temporal and parietal components of a distributed system for word meaning (semantic) representations. Speech repetition requires the pathway designated A in the figure, as well as more anterior parietal and frontal regions [not shown in colour] that support articulatory preparation and execution. Spoken word comprehension involves the pathway marked B in the figure, which maps perceived phoneme sequences to word concepts. Communicative speech production, in which the speaker retrieves words to express concepts, requires the pathway marked C, which maps concept representations onto phonologic representations. Pathway D indicates a direct mapping from visual word forms to phonologic representations, required for reading aloud’ (Figure and text taken from Binder (2015) [54])
Another common observation in functional imaging studies is that classic language regions are involved in multiple functions and, in the case of Broca’s area, also in different cognitive domains [59].6 In this respect, Broca’s area differs from Wernicke’s area, as it shows up as a component of many different nonlinguistic functions: motor imagination and preparation, music, visuospatial recognition, working memory and executive control (i.e. the organization of action and thought) [65–68]. Wernicke’s various subfunctions are basically all linguistic in nature [54, 55]. The fact that brain areas participate in multiple different functions implies that sets of brain areas can be temporarily bound together to perform a specific function [69]. Such a view would be consistent with the fact that different language tasks generally result in different brain activity maps (see Fig. 8.10 for an example). Conjunction analyses make use of this principle, hypothesizing that areas that are activated by different tasks play a more crucial role for the particular function that they (broadly) target [72]. Such an approach has been advocated for presurgical planning, in order to differentiate between areas that are supportive for a particular function and those that are critically needed or essential for normal performance [73].
