Structural magnetic resonance imaging

Structural MRI (sMRI) exploits the physical properties of atoms in the body to create high-quality 3D images of the various tissues so that their appearance and morphology can be examined. Visualisation is made possible because each type of tissue (e.g. bone, grey matter, white matter, cerebrospinal fluid) reacts differently to magnetic and radio wave frequencies. Although this technique is based on electromagnetic radiation, it does not involve exposure to ions or X-rays, which makes it relatively safe and allows for multiple exposures (Reference Goldstein, Price, Dougherty, Rauch and RosenbaumGoldstein 2004). SMRI uses a scanner containing a powerful static magnet. The strength of the magnet is measured in tesla (T) and it directly affects the potential definition and image quality. At present, most magnets in clinical use are 1.5 T or 3 T, but there are a few 7 T magnets in specialist centres. Interestingly, although the human body is made of multiple elements, it is the hydrogen in it that is exploited by MRI to generate the image. This is possible as hydrogen has high levels of both density and magnetic susceptibility (Reference Goldstein, Price, Dougherty, Rauch and RosenbaumGoldstein 2004).

Functional magnetic resonance imaging

Functional MRI (fMRI) employs the same magnet device, differences in tissue properties and physics used in sMRI to derive brain images (Fig. 1). However, the goal of fMRI is to visualise brain activity and so spatial resolution is sacrificed to the benefit of temporal resolution. To get an idea of the difference in image quality between the two techniques it should suffice to say that to acquire a scan of the entire brain takes in the order of 6 minutes for sMRI and 3 seconds for fMRI. Crucially, fMRI is not at sufficient temporal resolution to directly measure the neuronal signal and so it measures an indirect index of neuronal activity – the blood oxygen-level dependent (BOLD) contrast.

FIG 1 A high-quality fMRI scan from a 3 T scanner. Image courtesy of Dr O. O’Neil (own work); CC-BY-SA-3.0, via Wikimedia Commons.

Positron emission tomography and single-photon emission computed tomography

Positron emission tomography (PET) and single-photon emission computed tomography (SPECT) are used to investigate functional brain activity at the metabolic level. These techniques are based on the idea that it is possible to learn about a dynamic process by introducing a measurable tracer into a target system and then tracking this as it travels around the system. Indeed, both PET and SPECT involve the administration of a radiopharmaceutical tracer which, once absorbed, emits gamma rays that can be detected by the scanner and thus provide an image of the distribution of the tracer in the body over time (Reference Russell, Zelaya, Bressan, Fu, Senior and RussellRussell 2003). There are numerous criteria that the radiopharmaceutical needs to fulfil. It must not perturb the target system, but it must interact with the system in a predictable way, provide a quantifiable measure of concentration, show both high affinity and selectivity for the target, be safe for intravenous injection, be non-physiologically active at low doses and efficiently cross the blood–brain barrier. It is clear that the development of tracers for human use is not a quick or easy task.

Electroencephalography and magnetoencephalography

Electroencephalography (EEG) and magnetoencephalography (MEG) provide the most direct non-invasive measure of brain activity. Both tools detect activity associated with postsynaptic potentials, but EEG records the electrical fields of neurons, whereas MEG records their magnetic fields. In both techniques, the measurement comes from the scalp surface, via electrodes attached to a cap for EEG or via a complex system of magnetometers in a helmet for MEG (Fig. 2). As the signal generated by a single neuron is too weak to be detected at the scalp level, EEG and MEG can only pick up the waves coming from multiple neurons aligned in the same direction and with a synchronous flow. Thus, it is assumed that the main source of the signal comes from pyramidal cells along the cortex, as these are the ones best fitting these criteria (Reference Lopes da Silva, Mulert and LemieuxLopes da Silva 2010).

FIG 2 Magnetoencephalography (MEG) scanner with a person doing a task. Image courtesy of the National Institute on Aging/National Institutes of Health.

The main benefit of EEG and MEG is the great temporal resolution that they provide, recording at a very high sampling rate (up to 10 000–20 000 Hz or data points per second). However, the spatial resolution is quite limited, especially in EEG, where the thickness of the skull can easily smear the regional cortical signal and spread it across a broad area. This introduces a level of uncertainty into the estimation of the spatial sources of the signal. On the positive side, these procedures are better tolerated, less invasive, confer less risk and have fewer contraindications. MEG has a slightly better spatial resolution, but EEG can detect signals from deeper parts of the brain (Reference Goldenholtz, Ahlfors and HamalainenGoldenholtz 2009). Consequently, neither is always preferable over the other and some evidence suggests that a combined use gives the best results (Reference Goldenholtz, Ahlfors and HamalainenGoldenholtz 2009). However, as MEG is more expensive and requires sophisticated shielding, it is usually found in only a few specialist centres, whereas EEG is far more commonly used in clinical settings.

One of the most common clinical uses of EEG is in the diagnosis and treatment of epilepsy, as a way of measuring excessive or abnormal brain activity underlying the disorder. For example, it is routinely used to confirm the diagnosis of epilepsy emerging from clinical assessments and records. EEG can be indispensable in classifying seizure type and thus diagnosis, as sometimes these are based entirely on the electrical profile of the seizures (Reference Alarcon, Alarcon and ValentinAlarcon 2012). Also, EEG is often used to decide whether patients need surgery, for example by localising seizure foci, and to monitor the condition, for example by establishing the effects of anti-epileptic medication on seizure frequency (Reference Alarcon, Alarcon and ValentinAlarcon 2012).

In research, both tools are used to explore the neural pathophysiology underlying various mental disorders by providing detailed information about the temporal characteristics of relevant neural processes.

Other techniques

The techniques mentioned here are not the full story: more are available and some of these are outlined in Box 2.

Schizophrenia

One benefit of using neuroimaging in psychiatric research is its role in testing cognitive models of illness, especially using fMRI, which allows study of the brain while engaged in cognitive tasks. This is especially important for conditions such as schizophrenia, whose underlying pathophysiology is still unknown. For instance, a series of fMRI studies has helped to elucidate the formation of hallucinations in schizophrenia, demonstrating that both speech generation areas and speech perception areas are active during auditory hallucinations (Reference Shergill, Brammer and WilliamsShergill 2000) and that mechanistic treatments such as transcranial direct current stimulation (tDCS) or repetitive transcranial magnetic stimulation (rTMS) can be used to modulate brain activity in these regions and improve symptoms (Reference Moseley, Fernyhough and EllisonMoseley 2013). This supports a broader model that hallucinations reflect a misattribution of self-generated actions to others due to incapacity to differentiate between sensory signals deriving from external sources and those deriving from internal actions such as inner speech or thoughts (Reference Fletcher and FrithFletcher 2009). In support of this, research has found reduced BOLD activation in the somatosensory cortex of healthy controls when performing a self-generated (as opposed to externally generated) movement; however, people with schizophrenia lacked this attenuating mechanism and showed increased cortical activation in response to these self-generated movements (Reference Shergill, White and JoyceShergill 2014). Crucially, this activation was correlated with hallucination severity, thus corroborating the link between dysfunctions in self-monitoring and hallucinatory symptoms.

Applications of fMRI have also provided interesting insights into the psychological mechanisms underlying paranoid delusions, pointing towards lack of trust and reduced reward from social interactions as potential key elements. Using trust games, Reference Gromann, Heslenfeld and FettGromann and colleagues (2013) found that patients displayed reduced baseline trust of others. In addition, when playing with cooperative people, patients showed reduced caudate activity, which was inversely correlated with paranoia measures. Considering the established role of the caudate in reward-processing, and in positive trust especially, this has been interpreted as indicating a lack of perception of positive social interactions as rewarding. Last, important advances in schizophrenia diagnosis have been possible thanks to sMRI studies showing that specific patterns of abnormal grey matter can predict the illness with good levels of accuracy (e.g. Reference Nieuwenhuis, van Haren and Hulshoff PolNieuwenhuis 2012).

Depression

In the past 30 years, neuroimaging has been a major force driving our increased understanding of the neural pathophysiology of depression, identifying target structures for potential intervention. Neuroimaging data have identified the subcallosal cingulate gyrus (SCG) as a candidate structure for deep brain stimulation (DBS), which involves delivering electrical pulses via electrodes implanted in the brain. Furthermore, PET imaging has shown increased aberrant activity in the SCG of depressed people and of healthy individuals retrieving sad autobiographical memories. Also, a decrease in SCG activity has been observed after successful treatments employing various modalities (for a review see Reference Hamani, Mayberg and StoneHamani 2011). Thanks to such research, Mayberg and colleagues could successfully target the SCG of people with depression and show symptom remission and in 35% and amelioration in 60% of the sample (Reference Lozano, Mayberg and GiacobbeLozano 2008).

Importantly, imaging techniques such as PET and EEG can also be used to acquire further information about metabolic and electrical brain activity pre- and post-surgery. In their studies, Mayberg and colleagues found that DBS caused changes in the activity of cortical and limbic structures receiving SCG projections, providing insight into the specific region or network of structures modulating the beneficial effects of DBS (Reference Hamani, Mayberg and StoneHamani 2011).

Disordered gambling

Neuroimaging studies have influenced the categorisation of disordered gambling, recently moved from the category of impulse disorders to that of addiction in the latest version of the Diagnostic and Statistical Manual of Mental Disorders, DSM-5 (American Psychiatric Association 2013). This reclassification rests on various evidence of shared features between drug addiction and gambling disorder (Reference PotenzaPotenza 2006), including similar involvement of dopamine-mediated reward circuits, as demonstrated by neuroimaging.

Antisocial personality disorder and psychopathy

Antisocial personality disorder and psychopathy were largely viewed as mainly social constructs until neuroimaging data demonstrated biological differences in antisocial and psychopathic personalities (Reference Glenn and RaineGlenn 2014). Several neuroimaging studies have repeatedly shown functional and structural reductions in the brains of people with antisocial personality disorder or psychopathy which very well fit with their clinical profile. In particular, reduced volume and activity in the prefrontal cortex (PFC) (Reference Yang and RaineYang 2009) has been suggested to be associated with defects in social and moral judgements, emotional regulation and self-regulatory processes (for a review see Reference KoeningsKoenings 2012), while functional reductions in the amygdala (Reference Glenn, Raine and SchugGlenn 2009) might be linked to inappropriate emotion processing and fear conditioning (Reference Sehlmeyer, Schöning and ZwitserloodSehlmeyer 2009). Interestingly, some longitudinal neuroimaging studies have investigated whether key brain alterations could predict future antisocial behaviour. MRI is the ideal tool for this kind of study, as it is safe for repeated exposure. These preliminary studies conducted in prisoners and high-risk individuals showed reduced PFC activity and amygdala volume to be associated with offending behaviour 3 years later (reviewed in Reference Glenn and RaineGlenn 2014), suggesting that brain alterations may be causally involved in antisocial behaviour.

Neuroimaging has also been used to identify potential biomarkers for violence. Reference Raine, Lee and YangRaine and colleagues (2010) found that among people at risk of developing antisocial disorders (owing to their low socioeconomic status), those with a marker of disrupted limbic neurodevelopment, namely the presence of cavum septum pellicidum (CSP) as evidenced by sMRI, also showed increased violent and criminal behaviour. In line with evidence that CSP is associated with aggressive behaviour in animals, this suggests that predisposition to antisocial behaviour might be linked to early maldevelopment of limbic structures while also highlighting a potential role for CSP as a diagnostic marker.

Imaging genetics

There has been an explosion in the volume of genetic data available for understanding the links between genes, environment and mental disorder. Given that the impact of any genetic effect is likely to be via changes in the brain, there has been a drive to investigate the influence of genes on the structural and functional brain alterations underlying a given mental illness. One of the advantages of combining genetic and neural information is that genetics is more strongly related to the neural system underlying a given behaviour than to the behaviour itself (Reference Meyer-Lindenberg and TostMeyer-Lindenberg 2012). Thus, it is thought that imaging genetics will allow the detection of subtle differences that may not emerge with behavioural genetics. Much of this work is attempting to improve illness prediction and clinically useful disease stratification – identifying imaging and genetic biomarkers that can allow clinicians to target individuals at increased risk of developing the illness or that can be used to stratify patients into those who will benefit from one treatment rather than another.

For example, research has enabled us to link the increased risk of developing Alzheimer’s disease associated with e4 allele of the apolipoprotein E (APOE) gene to reductions in grey matter volume (Reference Shaw, Lerch and PruessnerShaw 2007) and functional connectivity (Reference Filippini, MacIntosh and HoughFilippini 2009) in young people, suggesting that brain imaging on people with this mutation might help early detection of brain abnormalities.

Another area of research is concerned with assessing the degree of heritability of a given brain structure or function known to be implicated in the pathophysiology of an illness. This is usually done by correlating brain similarity in individuals to various levels of relatedness. For example, if a given structural or functional characteristic is highly similar in monozygotic twins, less similar in dizygotic twins and even less so in siblings, then the genetic heritability of that characteristic is thought to be high. This method can be used to create a heritability map of brain areas activated during a given process, so as to show which areas are more dependent on genes than others.

Other studies are more interested in identifying which genetic variations contribute to key brain alterations. For example, Reference Hariri, Mattay and TessitoreHariri and colleagues (2002) linked specific allele variations in the promoter of the serotonin transporter gene to increased amygdala reactivity in response to anxiety-provoking stimuli, thus suggesting that genetic polymorphism might drive differences in anxiety and neuroticism traits. Once the link between genetic variation and brain activity has been established for a particular illness, one can use this information to identify those at risk of developing the illness and to inform the development of better treatments.

Resting-state fMRI

Until recently, most fMRI studies have focused on linking a given trait or behaviour to activation within a specific brain area, following a fairly traditional neurological lesion model. However, there is an increasing awareness that cognitive processes are largely supported by brain networks that operate across different brain regions. In line with this, resting-state study designs are particularly useful for acquiring information about the functional connectivity between brain regions.

Resting-state fMRI studies differ in a fundamental manner from routine fMRI studies in that the neuroimaging data are acquired while the person is at rest, i.e. not performing any explicit task. The main assumption behind this technique is that in functionally related brain regions, synchronised increases and decreases in blood flow signal neuronal activity. Thus, functional connectivity can best be estimated by measuring the correlation of low-frequency fluctuations between all the different brain regions, identifying several core networks that are correlated while at rest. Crucially, it has been shown that the same networks that are active in the resting-state mode are also active during specific tasks, thus giving insight into the cognitive functions underlying the identified networks. For instance, resting-state studies identified the so-called ‘default mode network’ as an area that is largely activated at rest, but strongly deactivated during goal-oriented activities. This enabled researchers to hypothesise and confirm its role in mind-wandering, spontaneous thinking and creativity (Reference Buckner, Andrews-Hanna and SchacterBuckner 2008).

In addition to their use in the study of functional connectivity, resting-state studies are also useful whenever active participation in a task is difficult (e.g. when testing children or sedated patients) or when failure to match performances between groups would significantly affect interpretation of data.

Machine learning

Computational approaches using machine learning have been applied for many years to problems in classification. However, only relatively recently have they been used in the classification of neuroimaging data to answer new questions not addressable using standard statistical approaches. Machine learning refers to the use of algorithms capable of automatically learning the statistical regularities in empirical data and using that learning to make predictions or decisions. In neuroimaging, this means using a subset of available sMRI or fMRI data to train an algorithm to learn data properties, typically in order to differentiate two groups of individuals, so that this algorithm can be applied to novel, unselected data.

Machine learning has two main advantages over standard approaches. First, it allows predictions to be made at the individual level. This is extremely clinically relevant, as the psychiatrist’s ultimate goal is to make a diagnosis, prognosis or treatment decision from the brain imaging data of the patient in question. Second, machine learning allows use of a wider range of data, including spatial correlation data, to examine the interactions between networks.

Most applications of machine learning in clinical research have aimed at identifying brain bio markers for the improvement of diagnosis and treatment response prediction. One of the most successful applications has been in Alzheimer’s disease: machine learning can diagnose the disease from sMRI scans with 90% accuracy (Reference Kloppel, Stonnington and ChuKloppel 2008a) and distinguish it from other forms of dementia more accurately than radiologists can (Reference Kloppel, Stonnington and BarnesKloppel 2008b). This is an invaluable contribution to psychiatry, as only about 75% of diagnoses of Alzheimer’s disease made by clinicians are confirmed by post-mortem studies.

Another disorder that has been difficult to diagnose is autism, as the optimal diagnostic tool is time-consuming and requires the gathering of retrospective information. Studies show that machine learning applied to MRI data on cortical thickness can predict autism with 90% accuracy (Reference Ecker, Marquand and Mourão-MirandaEcker 2010). Important achievements have also been made in major depressive disorder, where MRI at baseline is now able to predict treatment response over time (Reference Costafreda, Chu and AshburnerCostafreda 2009).