Introduction

Since Emil Kraepelin’s conceptualization of endogenous psychoses as two categories, dementia praecox and manic depression, the separation between primary psychotic disorders and primary affective disorders has been much debated.1,2,3,4 There is substantial evidence that these illnesses share genetic risk factors and overlap in clinical presentation and treatments, suggesting these clinical phenotypes to be on a disease continuum or even differential expressions of the same illness.5 Contemporary models attempt to move towards classification of psychiatric disorders based on etiologic and pathophysiologic processes, integrating complex relationships between genetics, physiology, and behavior.6, 7

Since the initial description of the principles of magnetic resonance imaging (MRI) by Paul Lauterbur in 19738 and the first human MRI scan in 1977,9 much progress has been made in the field of neuroimaging. A variety of contemporary non-invasive sequences are now available to aid structural, functional, and neurochemical characterization of the human brain, providing new opportunities to shed light on biological and pathological processes in vivo.10

Conventional, T1-weighted, structural MRI images provide static anatomical information with excellent detail and strong gray/white matter contrast. Manual tracing methods examine brain morphology of pre-defined regions of interest (ROI) such as the hippocampus, but are quite labor intensive. Computational advances offer semi-automated and automated alternatives to manual tracings. One of the most popular automated techniques is voxel-based morphometry (VBM11), which quantifies brain structure at the voxel level, with a typical resolution of 1 mm3. Another frequently used automated structural imaging analysis technique is Free Surfer,12 which also examines cortical thickness, cortical gyrification, and shapes of sub-cortical structures. Diffusion weighted imaging techniques, such as diffusion tensor imaging (DTI), map three dimensional motion of water as a function of spatial location to describe anatomy of anisotropic tissues such as white mater.13 The most widely reported diffusion tensor measures are fractional anisotropy (FA) and mean diffusivity (MD) which describe complementary information on structural white matter integrity. Tract based spatial statistics (TBSS) is an automated method describing diffusion metrics from white matter in the entire brain,14 whereas fiber tracking quantitatively assess the microstructure of a specific white matter tract.15 Fiber reconstruction methods are either deterministic or probabilistic; deterministic methods follow the primary eigenvector from voxel to voxel in three dimensions, whereas probabilistic methods incorporate expected uncertainty into the algorithm.16

Functional MRI (fMRI) imaging provides dynamic physiological information by measuring the blood oxygen level dependent (BOLD) signal.17 Task fMRI characterizes brain activity while subjects perform cognitive tasks by measuring changes in the BOLD signal in different areas of the brain. In block designs different conditions are alternated in blocks, where the condition assessing the cognitive process of interest is alternated with a control condition, which makes in the change in the fMRI signal in response to the stimulus additive.18 Block designs offer considerable statistical power, but are prone to signal drift and run the risk of having subjects become aware of the order of events. On the other hand, event related designs are more reflective of the real world, but come at the expense of statistical power.19 Alternatively, resting state fMRI measures the temporal covariance of low frequency fluctuations of the BOLD signal across spatially disparate areas while no explicit task is performed, in an effort to assess the brain’s intrinsic functional organization.20 Seed based analysis is a popular hypothesis-driven method to visualize resting state networks.21 In this case, an area of interest is defined from which correlations in BOLD fluctuations with all other voxels in the brain are calculated, allowing to examine connectivity of a specific ROI. In contrast, independent component analysis (ICA) is a data-driven method that decomposes the multivariate signal across the brain into statistically independent components (either spatially or temporally) reflecting resting state networks.22, 23

Magnetic resonance spectroscopy (MRS) measures chemical composition of tissues, energy metabolism, and neurotransmitter levels in vivo.24 The most common type of spectroscopy leverages the signal produced by protons located in the molecules of living tissue to quantify different metabolites (1H-MRS). Metabolites that can be measured include N-acetyl-aspartate (NAA), a putative marker of neuronal integrity, choline (Cho), a marker of cellular turnover and cell membrane breakdown, creatine (Cr), a signal that is related to phosphate metabolism, and the amino acids glutamate (Glu), glutamine (Gln), often expressed as Glx (glutamate + glutamine), and ɣ-amino-butyric acid (GABA).25 Other spectroscopy techniques include 31Phosphorus spectroscopy, which provides a wide range of information on energy metabolism, and 13Carbon spectroscopy which use cerebral glucose metabolism to assess glutamine synthesis and glutamatergic transmission.

These complementary imaging techniques have been applied to better delineate the neurobiology of psychiatric disorders in vivo. Structural gray and white matter deficits, as well as neurometabolite alterations, and BOLD signal abnormalities during task and at rest have been found in schizophrenia and bipolar disorder (BD). However, it remains unclear to what extent neural signatures are converging or distinct.

Here, we conducted a systematic review of case–control studies contrasting structural, functional, and neurochemical abnormalities in BD, schizoaffective disorder, schizophrenia, and healthy controls conducted in the past 10 years with the objective to summarize progress made in the quest to better delineate pathophysiological patterns across the psychosis spectrum. Where relevant, selected older publications considered key developments within the field and meta-analyses using quantitative methods to synthesize the literature are also included.

Results

Study identification

Figure 1 describes outcomes at each level of our study identification process. Of the 394 potentially relevant articles, we included 50 structural, 29 functional, seven MRS, and eight combined imaging and genetic studies in this systematic review.

Fig. 1
figure 1

Process of study selection

Full size image

Structural MRI studies

Gray matter structural MRI

Gray matter has been an early focus of neuroimaging research in psychiatric illnesses. Using anatomical likelihood estimation (ALE), a meta-analysis of 37 studies demonstrated extensive gray matter deficits in frontal, limbic and subcortical deficits in schizophrenia when compared to healthy controls.26 Similarly gray matter reductions were found in anterior cingulate and insula in an ALE meta- analysis of patients with BD when compared to healthy controls.27 A recent analyses attempting to quantitatively, albeit not directly, contrast gray matter deficits in a number of psychiatric disorders, including schizophrenia and BD, using meta-analytic techniques found shared decrease in the dorsal anterior cingulate and anterior insula, areas of the brain that are attributed to the salience networks.28

Forty studies explicitly examined differences in gray matter across the psychosis spectrum since 2005, but only seventeen of those were conducted at magnetic field strengths of 3T. All but one study included subjects who were medicated at the time of scanning, and only a minority of studies focused on first episode patients (Table 1). Cortical gray matter volume loss appears widespread in schizophrenia, but less extensive29,30,31 or even absent32, 33 in BD. It has been suggested as an intermediate phenotype across disease categories, possibly reflecting lifetime psychosis burden, with patients diagnosed with schizophrenia and schizoaffective disorder demonstrating extensive neocortical and subcortical gray matter reductions, and smaller reductions limited to frontotemporal regions in BD.29 Further supporting this concept is a recent study by Song et al. who examined gray matter volumes in unmedicated patients with schizophrenia and BD and reported a negative correlation between severity of delusions and frontal gray matter volumes as well as extent of hallucinations and right uncus gray matter volume across diagnostic groups.34 Examining subcortical areas of the brain, studies consistently suggest that hippocampal volume reduction may be a feature that is shared across the psychosis spectrum,35,36,37, with a majority of studies reporting that volume loss is greater in schizophrenia compared to BD,38 even when examining individual hippocampal subfields.35, 39 Similarly, early studies suggest thalamus volume reductions to be present across disease categories,36, 40 but later reports with larger sample sizes detected this feature only in schizophrenia.29, 41 Findings are more inconsistent in regards to amygdala volume, which has been reported to be unaffected in both diagnostic groups,42, 43 decreased only in schizophrenia,44 more prominently decreased in schizophrenia compared than in BD45 and vice versa.37 Reports on basal ganglia volumes are also conflicting where some find volume increase46, 47 or decrease48, 49 that is shared across illnesses, abnormalities in schizophrenia but not BD,34, 36 or lack of abnormalities in either diagnostic group (except for the nucleus accumbens).50 Those contrasting reports likely reflect the heterogeneity across studies in regards to patient characteristics, medication exposure, data acquisition, and data analysis methods.

Table 1 Studies examining gray matter
Full size table

Significant cortical and subcortical volume loss that resembles the chronic illness stage is already reported in first episode patients. While many of the areas with gray matter loss appear to overlap across diagnostic groups, several reports suggest greater abnormalities in first episode schizophrenia compared to first episode BD, both in terms of volume loss45, 51 and spatial extent.40, 52 It is perhaps not surprising that Pina-Camacho et al. reported that age at first onset of psychosis modulated structural abnormalities in a nonlinear and diagnosis dependent manner. Specifically, they report that patients with an earlier onset of a schizophrenia spectrum disorder had the most significant ventricular and basal ganglia enlargement along with the greatest fronto-temporal cortical volume and thickness deficits among diagnostic groups, with affective disorder patients having less extensive cortical deficits that were again more prominent in those with younger age of onset of psychosis.53 However, none of the studies examined medication naïve patients, making it impossible to definitively conclude that observed group differences in gray matter abnormalities are due to intrinsic differences across the diagnostic spectrum rather than secondary to differential exposure to psychotropic medications.

White matter structural MRI

White matter abnormalities have been reported to be widespread in both schizophrenia and BD. An activation likelihood estimation meta-analysis showed decreased FA in first episode schizophrenia compared to healthy controls across the commissural, association, and projection tracts, with main involvement of the inferior fronto-occipital fasciculus, inferior longitudinal fasciculus, cingulum bundle, and corpus callosum.54 Similarly, a meta-analysis of fifteen DTI studies in BD reported decreased FA in all types of tracts when compared to healthy controls, with the most robust decreases in the inferior fronto-occipital fasciculus.55 Findings suggest a shared spatial distribution of white matter integrity deficits across the illness spectrum, but allow no inference on the comparability in magnitude of abnormalities.

Eleven studies contrasted white matter integrity in schizophrenia and BD; all but one included patients who were medicated at the time of assessment, and only two focused on first episode patients (Table 2). It is striking that the majority of studies conducted in chronic patients on psychotropic medications show decreases in white matter integrity that do not appear different across diagnostic groups. Region of interest analyses show shared FA decreases in the uncinate fasciculus, corona radiata, anterior limb of the internal capsule, anterior and posterior thalamic radiation, and corpus callosum.56,57,58,59,60 This is also corroborated by Skudlarski and colleagues who conducted the largest study thus far and found a close agreement between spatial distributions and magnitudes in FA reductions assessed with whole brain TBSS across diagnostic groups. Interestingly, they reported higher variance in patients with psychotic BD, suggestive of greater heterogeneity in white matter integrity abnormalities compared to patients with schizophrenia.61 Heterogeneity may also explain discrepancies with Anderson and colleagues who found FA reductions in temporal and occipital white matter in schizophrenia but not bipolar I disorder,62 and Knoechel and colleagues who report that the magnitude of white matter integrity abnormalities in the left cingulum and right uncinate fasciculus is greater in schizophrenia than BD.38

Table 2 Studies examining white matter
Full size table

In first episode patients, no white matter volume abnormalities or no differences across diagnostic groups were observed63. The only study to date examining white matter microstructure in medication-naïve and minimally treated patients found a shared increase in MD in a large number of white matter tracts across diagnostic categories, but showed that patients with first episode BD had decreased FA in the cingulum, internal capsule, and posterior brain regions that was not evident in first episode schizophrenia.64

Functional MRI studies

Task functional MRI

A total of thirteen studies, all in chronically medicated patients, have directly contrasted functional activation patterns in BD and schizophrenia using various tasks (Table 3). In working memory tasks, studies largely observe engagement of the same brain networks in healthy controls and patients across the psychosis spectrum, but find alterations in activation patterns within those networks. A graded pattern of group differences in the amplitude of the BOLD signal has been reported in several studies, with the greatest alteration typically reported in schizophrenia, and more subtle or lack of abnormalities in BD.65,66,67 Findings are more inconsistent, with both hypo- and hyper-activation reported, which may be explained by different, but overlapping, inverted u-shaped curves of activation depending on task difficulty68 across the psychosis spectrum, with less abnormal patterns of activation in BD compared to schizophrenia. Alternatively, it is possible that these patterns of differential activation could represent compensatory processes or secondary effects of primary changes in signal processing. Greater task related alterations in schizophrenia compared to BD were also reported in a verbal fluency task,69 but not a sentence completion task,70 or emotionally salient memory tasks,71, 72 the latter appearing more altered in BD than schizophrenia. Furthermore, activation during reward anticipation in the ventral striatum appears decreased in schizophrenia, but not patients with BD in a manic state.73 The authors speculated that striatal dopamine dysfunction, which could be clinically expressed as anhedonia, may be underlying their finding.

Table 3 Studies examining task activation
Full size table

Resting state functional MRI

Investigation of functional connectivity at rest has become increasingly popular, in part because task performance differences between groups need not be accounted for in this paradigm.74 An activation likelihood estimation meta-analysis of whole brain resting state studies in schizophrenia suggests decreased activity in the medial prefrontal cortex, left hippocampus, posterior cingulate cortex and precuneus (all areas of the brain that are typically conceptualized as part of the default mode network75), as well as increased activity in the lingual gyrus.76 A recent review attempting to reconcile methodological differences in schizophrenia studies suggested increased functional connectivity to be a replicated finding.77 In a qualitative systematic review Vargas and colleagues reported aberrant resting state connectivity in between frontal and meso-limibic areas in BD when compared to controls.78

All of the sixteen resting state studies comparing connectivity across diagnostic groups included here were conducted in patients who were medicated at the time of scanning (Table 4). The default mode network, a large scale brain network that is more active at rest and has been implicated in self-referential thinking, is perhaps the most widely studied. An early report in acutely ill patients with schizophrenia and BD identified the medial prefrontal cortex as major locus of shared abnormality, with BD being characterized by reduced default mode network connectivity to the hippocampus and fusiform gyrus as well as increased connectivity with the primary visual cortex, and schizophrenia being characterized by abnormal recruitment of the frontal polar cortex and the basal ganglia.79 The largest study to date examining default mode network connectivity with ICA reported connectivity reductions in the medial prefrontal cortex, anterior cingulate cortex, posterior cingulate cortex and precuneus across the psychosis spectrum, but also found that selective nodes within the network appear to be differentially affected in schizophrenia and BD.80 The same group also reported aberrant connectivity between the default mode and a fronto-occipital network as shared illness feature in schizophrenia and BD, whereas increased connectivity between fronto-temporal and mesolimbic regions was only evident in BD, and decreased connectivity between sensory-motor and mesolimbic areas was limited to schizophrenia.81 Others report within and between network connectivity decreases within a cingulo-opercular network and between a cingulo-opercular and cerebellar network that are shared across illnesses, decreased connectivity between the cingulo-opercular and salience network in BD only, and decreased connectivity between cingulo-opercular and fronto-parietal network in schizophrenia only. Notably, default mode network connectivity was not reported abnormal.82 Similarly, Baker and colleagues reported resting state connectivity disruptions of cortical association networks, preferentially the frontoparietal control network, but not default mode network abnormalities, in schizophrenia and BD.83

Table 4 Resting state studies
Full size table

Resting state studies of non-neocortical structures revealed distinct patterns of thalamic84 and amygdalar85 dysconnectivity in schizophrenia and BD, but no differences across diagnostic groups in hippocampal38, 86 connectivity. Examining connectivity across the entire brain with global brain connectivity, a measure that computes connectivity strength of every region of the brain with every other region of the brain, was reported to be lower in schizophrenia than in healthy controls, whereas patients with BD had intermediate global connectivity strength that was significantly different from both patients with schizophrenia and healthy controls.87

Magnetic resonance spectroscopy studies

There is substantial evidence that neurometabolite levels are altered in both schizophrenia and BD. A meta-analysis pooling data from 146 studies suggests decreases in NAA in the frontal lobe, hippocampus, thalamus, and basal ganglia in schizophrenia, but only in the basal ganglia and frontal lobe in BD.88 Another meta-analysis summarizing findings of glutamatergic abnormalities across 28 studies in schizophrenia revealed a decrease in medial frontal glutamate compared with healthy controls,89 but the majority of studies were conducted in medicated patients. Contrastingly, several reports do suggest an elevation of glutamatergic indices in unmedicated patients with schizophrenia in the medial prefrontal cortex, striatum, and hippocampus.90,91,92,93 A smaller meta-analysis in BD including nine studies measuring Glx (a combination of glutamate and glutamine) across different areas of the brain, suggested that this metabolite may be higher in patients with BD compared to controls, irrespective of medication status.94 Taken together, it appears that some of the neurometabolite alterations, specifically decreased NAA in the frontal cortex and basal ganglia may be shared across the illness spectrum, whereas others may not.

However, studies directly contrasting neurometabolites in BD and schizophrenia are sparse. All of these studies used single voxel1 H-MRS placed in cortical areas of the brain and were conducted in chronically ill patients who were medicated at the time of assessment, most commonly with very small sample sizes (Table 5). Molina and colleagues were the first to report that NAA/Cr decreases in the left, but not the right dorsolateral prefrontal cortex were greater in schizophrenia than in BD, with no Cho/Cr abnormalities appreciated in either group.95 Findings were partially replicated by Kalayci and colleagues, who reported a universal decrease in left and right dorsolateral prefrontal cortex NAA/Cr across diagnostic groups, but a decrease in Cho/Cr only in BD and schizoaffective disorder, but not schizophrenia when compared to controls.96 Anterior cingulate cortex metabolite measurements at 1.5 Tesla were suggestive of elevations of Cho/Cr in schizophrenia only without detectable abnormalities in NAA/Cr across groups.97 However, because of a later report of decreased Cr levels in acutely ill patients with schizophrenia but not BD, it is unclear if findings are attributable to Cr alterations rather than Cho or NAA changes.98 In the left Heschl’s gyrus, Glu, NAA, and inositol levels were found to be decreased in BD, but not in schizophrenia when compared to healthy controls, while no metabolite abnormalities in either diagnostic group were detected in the right Heschl’s gyrus, suggesting a lateralized abnormality in the dominant hemisphere.99 Only one study to date has examined GABA, with reports of perisylvian GABA elevations in patients with schizophrenia, but not those with BD, when compared to healthy controls.100

Table 5 Studies examining magnetic resonance spectroscopy (MRS)
Full size table

Imaging-genetic studies

Eight studies, all in medicated patients, have examined relationships between imaging and genetic markers across the illness spectrum (Table 6). In verbal fluency tasks, significant diagnosis by genotype interactions with task related activations were observed for Neuregulin 1,101 disrupted in Schizophrenia Gene 1 (DISC 1102), and the d-amino acid oxidase (See ref. 103) gene. Interestingly, a modest relationship between white matter volume and a number of schizophrenia risk genes was found across the psychosis spectrum and in healthy controls, suggesting that cumulative genetic risks may help explain the extent of observed white matter alterations,104 but other risk genes did not show such relationships.105

Table 6 Imaging genetic studies
Full size table

Discussion

In this systematic review, we sought to summarize converging and distinct neural signatures in schizophrenia and BD. Structural neuroimaging studies suggest white matter integrity deficits that are consistent, both in magnitude and spatial extent, across the psychosis spectrum, while gray matter reductions, especially those that are cortical, appear more widespread in schizophrenia compared to BD. Similarly, spectroscopy studies in cortical gray matter report evidence of decreased neuronal integrity in both disorders, but not enough data exists to draw firm conclusions as to differences in magnitudes and spatial distribution between illnesses. On a functional level, findings are inconsistent, possibly because of small sample sizes in many of the studies. Functional MRI studies using task paradigms typically report engagement of the same brain networks in healthy controls and patients across the psychosis spectrum, but find differential extent of alterations in magnitude of task related activation between illnesses depending on the task paradigm. The larger resting state connectivity studies are inconsistent as to abnormalities in the default mode network, but it appears that decreased fronto-parietal network connectivity may be a shared feature across the psychosis spectrum. The very limited imaging-genetic literature suggests a relationship between psychosis risk genes and brain structure, and possible gene by diagnosis interaction effects on functional imaging markers.

Our work needs to be considered in context of several limitations. We performed a systematic review, but did not perform quantitative assessments using meta-analytic techniques which could be informative in future work. We did not do separate analyses in schizophreniform disorder or schizoaffective disorder due to a lack of studies investigating these as a distinct disease category (schizoaffective disorder is most commonly grouped under schizophrenia). We decided not to include ultra-high risk subjects as conversion rates to psychotic illness is reported to be 30% or less,106 and psychotic depression due to the paucity of studies including this disease category.

It is important to note that the vast majority of studies have been conducted in chronically ill, medicated patients, which precludes us from being able to disentangle intrinsic illness characteristics from changes attributable to disease progression and exposure to psychotropic medications. While the existing literature suggests some shared and some distinct neural markers in across the psychosis spectrum, it will be imperative to conduct large, well designed, multi-modal neuroimaging studies in medication-naïve first episode patients that will be followed longitudinally over the course of their illness in an effort to advance our understanding of disease mechanisms and to resolve the illness dichotomy vs. illness spectrum debate. But because this type of studies are notoriously difficult to conduct, and sample sizes are typically fewer than 50 subjects,107,108,109,110,111,112 a potential alternative strategy may be to obtain very large, multi-site datasets and attempt to mitigate medication confounds with statistical adjustments.

Methods

Eligibility criteria

Studies were included if they presented original data published between January 2005 and December 2016 (last search December 9th 2016), compared individuals with BD and schizophrenia/schizoaffective disorder and healthy controls. Studies were not included when the healthy control group was genetically related to the patient groups. Studies published in languages other than English, post mortem studies, non-human studies, and review articles were excluded. We only included trials with ten or more healthy subjects, aged 18 or older. Studies expressively including subjects with comorbid substance use disorders, neurological or genetic diseases, or intellectual disabilities were not considered. When a single study was published in several articles, the article reporting the largest group was used. Articles that did not explicitly compare imaging findings between diagnostic groups were excluded as well.

Literature search

BB and NVK performed a literature search in PubMed including subjects across the psychosis spectrum using the following key words: (Schizophrenia AND bipolar disorder) AND (gray matter OR morphometry OR VBM OR volume OR white matter OR DTI OR magnetic resonance spectroscopy OR MRS OR functional MRI OR resting state). The reference lists of included studies, as well as relevant meta-analyses were inspected for additional eligible publications.

Study selection

After removal of duplicate articles, BB and NVK screened titles and abstracts retrieved from the search and selected potentially eligible studies for full text review. Both authors applied eligibility criteria, and a list of eligible full text articles was developed through consensus. Full text articles were then downloaded or requested from the university library and assessed for eligibility. Figure 1 describes the study selection process and outcome.

Data extraction

We extracted the following data from each study: name of first author, year of publication, number of participants per diagnostic category, illness duration, mood state, use of psychotropic medications, data acquisition parameters, magnetic field strength, data processing parameters, main study outcomes.