Epilepsy, a serious health condition characterized by recurrent and often disabling seizures (1–3), is the fourth most common neurological disorder (4), with prevalence ranging from 0.5 to 1% around the world (5, 6). Despite over a century of drug development (7), about one third of these cases are medically intractable (8–10). This is perhaps in part due to the approach often taken, agnostic to seizure microcircuitry and network physiology. Surgical interventions such as ablation (11, 12), open resection (13, 14), and electrical stimulation (15, 16) are sometimes options for pharmacoresistant epilepsy. However, none of these procedures render all patients seizure free. Additionally, they can result in off-target effects, and are generally unavailable in developing countries. Thus, a significant need exists for more targeted and effective therapies (17, 18). A deeper understanding of the role of individual neurons, neuronal ensembles, and neural networks in initiating, propagating and terminating epileptic discharges underlying seizures would facilitate such advances.

Currently, the bulk of our knowledge of seizure dynamics comes from macroelectrode population electrophysiology, with electroencephalography (EEG; Figure 1A) having demonstrated that seizure activity can spread from a focal brain region in a diffuse yet stereotyped network (19). However, this approach fails to capture the complex underlying microcircuit dynamics, many permutations of which can result in the same recorded signal (18). To characterize seizure dissemination between individual neurons, single unit recordings have been performed in humans and animals using microelectrode arrays (MEAs). These recordings have shown that firing and termination of firing of neurons recruited during seizures is highly synchronous and stereotyped (20, 21), where ictal wavefronts propagate with similar directionality across seizures and interictal spikes propagate in antiparallel fashion (22). Additionally, electrophysiology combined with GABAergic pharmacology has been used to find evidence of inhibitory networks restraining seizure activity, including surround and feedforward inhibition (23, 24). However, these recordings have limited recording density, typically at a single cortical depth, due to their planar configuration of sparsely arranged contacts. Furthermore, it is generally impossible to know the subtype of neuron being recorded [for exceptions see (25, 26)]. Moreover, it is challenging to determine if the activity being recorded emanates from the same neuron between separate recording sessions.

Functional optical imaging in animal models (Figure 1B) circumvents many of the limitations of electrophysiology and allows for observation of activity of substantially more neurons. Calcium (27–30) or voltage sensitive dyes (31, 32) were originally used for this purpose, but were rarely cell-type selective, and disadvantaged by significant photobleaching, poor intracellular retention and toxicity (33). The development of genetically encoded indicators, expressed either through viral transduction or through genetic model development, has enabled cell-type specific imaging. These fluorescent proteins undergo a conformational change to excitable states upon the binding of specific ions or small molecules (34, 35). The most commonly used indicators are genetically encoded calcium indicators (GECIs) that fluoresce in the presence of calcium (35–41). They can serve as proxies for neuronal firing, which is marked by an increase of intracellular calcium. In addition to calcium, a growing arsenal of genetically encoded indicators are being developed (34, 42), including for other ions (43–48), small molecules and neurotransmitters (49–58). Furthermore, genetically encoded voltage indicators (GEVIs) have been developed, exploiting voltage-sensitive domains (59–63).

For imaging, these indicators can be excited with either a single photon (1P) in the visible range or two (or more) coincident photons in the infrared range. While 1P excitation can be achieved with many intravital microscope setups, including epifluorescence widefield, laser scanning confocal and light sheet, 2P necessitates the use of a femtosecond mode-locked laser for light delivery (64). Consequently, 2P imaging is limited to laser scanning, although a few light sheet uses exist (65). The temporal and spatial resolution of these methods varies greatly (17, 64, 66–68) and thus the selection of the method should be tailored to the question and indicator. 1P widefield imaging offers the largest fields of view with fastest temporal resolution (Figure 1C). However, for investigations looking at subcellular compartments or multiple cell populations of tens to hundreds of individual neurons, 2P laser scanning is often best. For questions related to brain-wide mesoscale networks, light sheet and 2P laser scanning in transparent zebrafish larva have been the methods most often used. All the above methods require head fixing or immobilizing the subject under an objective. Should the question necessitate a wider range of behavior or longer imaging session, head mounted microscopes [1P (69) and 2P (70, 71)] can be used in freely moving mice. For a more detailed review of epilepsy intravital imaging methodology see (66).

Since the first intravital optical imaging study using a genetically encoded indicator in a vertebrate seizure model nearly a decade ago (72) there have been over 50 additional such studies published. These have sought to examine the contributions of different cell types and neurotransmission to seizure dynamics, the reliability of initiation and propagation patterns, the network changes in synchrony and connectivity that occur at micro and macroscale, and the impact of this aberrant activity on normal function. This review will serve to summarize the major contributions these investigations have provided to our understanding of seizure physiology.

Seizures can be partitioned into several distinct phases, namely interictal, pre-ictal, ictal, and post-ictal (Figure 1A). While each phase’s distinct dynamics are of importance, it is the evolution of that activity and the transitions between phases that are often the focus of investigation. Therefore, we have organized our review in a complementary fashion. Table 1 lists the in vivo imaging studies included in this review, specifically those that used genetically encoded indicators to investigate epileptiform activity directly. Experiments where imaging was not in vivo (e.g., ex vivo slice), used dyes, or was only performed during non-epileptiform activity are not discussed.

Interictal spikes (IIS) are episodes of transient synchronous paroxysmal depolarization across ensembles of hyperexcitable neurons, classically observed as a spike wave discharge on EEG (73). Widefield imaging of the cortex has demonstrated that IIS begin as standing waves in local regions with limited contiguous spread (74), although some cortex-wide propagation and delayed recruitment of distal non-contiguous foci has been observed (75, 76). While IIS are often limited to or emanating from the purported epileptic focus, in instances of extrafocal origin, or non-contiguous spread, these loci often share homotopic connection with the ictal focus (74, 77). These patterns have been observed in both excitatory and inhibitory populations (77). The local and limited nature of interictal activity is corroborated by whole brain imaging in zebrafish (78–80).

To parse the individual cell activity patterns underlying IIS, laser scanning microscopy was employed. Imaging the zebrafish optic tectum revealed hypersynchronous recruitment of microensembles underlying the spatially confined interictal activity observed by previous studies with widefield microscopy (79, 81). In the pilocarpine chronic seizure mouse model, it was shown that inhibitory neurons are disproportionally active relative to pyramidal cells during IIS in CA1, consistent with a perisomatic inhibitory restraint occurring. Additionally, while it seems there are subpopulations of neurons consistently active together, their recruitment is varied across spikes (72). However, in a chronic intrahippocampal kainic acid (KA) mouse model, synchronous bursts of pyramidal cells have been observed, which could speak to model differences in ictogenic mechanisms (82). When imaging the dentate gyrus (DG) also following intrahippocampal KA, distinct microensembles of excitatory adult born granule cells (abGCs) were determined to overly drive IIS, albeit firing in a desynchronized manner. The specific ensembles recruited across IIS were varied. These are distinct from the microensembles that participate in sharp wave ripples, which were shown to be driven by both mature and abGCs, firing with greater synchrony. This suggests that decoupling of abGCs from mature GCs and subsequent reorganization into these desynchronized pathologic ensembles may contribute to an impairment in dentate gating, enabling ictogenesis (83). When training computational models on calcium data from this model, abGCs were most often identified as superhub neurons with high feedforward conductance, enhancing downstream excitation in the resulting epileptic networks (84). Neurons outside the hippocampus, but still part of this network have also been found to be involved in seizure regulation, such as the fasciola cinereum, a collateral intermediary nucleus connecting the entorhinal cortex to the DG (85). Gap junctions may be in part mediating the spread of IIS activity, particularly in the astrocyte syncytium, as blocking gap junctions significantly decreased the occurrence, duration and spread of IIS (79). However, while astrocytes in CA1 exhibit transient increases in calcium spontaneously during the interictal period, these have been observed asynchronous with IIS (86).

For those spikes occurring in the cortex there is a notable recruitment of pyramidal cells (87, 88), limited to layer II/III, in addition to inhibitory cells (89). While both parvalbumin (PV) and somatostatin (SST) cells demonstrate activity during IIS, PV cells were predominantly recruited with a higher degree of inter- (with pyramidal cells) and intrapopulation synchrony, while SST cells demonstrated asynchronous and delayed recruitment (89). Combined GEVI and GECI imaging in pyramidal cells revealed that there is little supra- or subthreshold activity propagating out of the focus during IIS (90). Studies utilizing neurotransmitter indicators were concordant with these IIS dynamics, revealing an increase in glutamate observed at the focus, which expanded centrifugally, and an increase in gamma-aminobutyric acid (GABA) observed extrafocally, which displayed slower and more persistent centripetal propagation, consistent with intact feedforward inhibitory surround limiting the IIS spread (51, 91). Indeed, the IIS may reflect this restraining mechanism (92). Importantly, the majority of these studies were performed in focal neocortical models, where a chemoconvulsant was intracortically injected to elicit epileptiform activity. Corroborating studies across other models could strengthen the generalizability of these findings.

Combined widefield GECI imaging with optical imaging of intrinsic hemodynamic signal to examine neurovascular coupling showed that during IIS there is an initial ‘epileptic’ dip in hemoglobin oxygenation, likely the result of vasodilation, followed by a period of hyperoxygenation, due to increased levels of total hemoglobin delivery, all of which is tightly spatially correlated with excitatory cell activation (76, 93).

In terms of functional impact, hippocampal place cell encoding is impaired in two chronic seizure models, with aberrant dendritic hyperexcitability (82) and aberrant firing and desynchronization (94) hypothesized as contributing mechanisms.

Taken together, intravital imaging at microcircuit resolution has revealed distinct cell type activity patterns specific to different anatomical regions during IIS, and speaks also to intra- vs. extrafocal differences (92). There is considerable variability in the specific neurons recruited, where different ensembles can be recruited across sequential IISs.

Interictal periods, by definition, occur between seizures and thus they have both pre-ictal and post-ictal phases, the exact boundaries of which are poorly defined. Thus, to examine the pre-ictal phase, we look at the interictal period from the perspective of the progression of dynamic changes as the brain state transitions to seizure.

An advantage of imaging is the ability to delineate activity as intra- or extrafocal. When recording from the focus in a mouse 4-aminopyridine (4-AP) model, recruitment of both excitatory and inhibitory neurons in local microensembles, akin to microseizures, was observed prior to ictal onset. As the brain progressed to seizure, ensemble activity increased across populations, with synchronization amongst excitatory cells (95). When examining dynamics at mesoscales in zebrafish whole brain, the mesencephalon/optic tectum, and occasionally the thalamus, emerged as a conserved region of hyperactivity (96–98) and seizure focus (78, 80, 81, 99–101). It was found that foci tended to have a higher proportion of excitatory cells than in the penumbra, perhaps contributing to their hyperexcitability (80).

When exploring cortical regions outside of the seizure onset zone prior to seizure invasion, feedforward inhibitory activity was observed, be it increased firing activity of inhibitory neurons (23, 95, 102, 103), or elevated GABA release during pre-ictal spikes relative to within the focus (91). Excitatory recruitment was still discernable, although possibly with some degree of suppression (95). As the tissue transitioned to a seizure state, there was a progressive increase in synchronization amongst pyramidal cells (89, 102), with a gradual breakdown in the inhibitory surround, witnessed as desynchronization in the PV cells activity (89, 102) and a decrease in released GABA during spiking (91). Inhibitory restraint weakening is also detected as progressively increasing bursts of subthreshold activity in pyramidal cells by GEVI imaging in the penumbra (90). As compared with neurons, astroglial networks displayed a more widespread elevation in calcium activity and synchrony pre-ictally, although their bursts of activity seemed to follow immediately after neuronal bursts during pre-ictal spikes (98).

Modeling based upon whole brain zebrafish mesoscale 1P light sheet imaging during PTZ induced seizures found that the tectum served as a networkwide hub. As the brain transitioned to seizure, there was a decrease in input to the tectum, enabling downstream network synchronization. The brain-wide recruitment was facilitated by increasing faster excitatory transmission and decreasing slower inhibitory transmission (104). Similar findings emerged from a model trained on 2P data with single-cell resolution collected from the same model. Specifically, they found that pre-ictal networks had enriched feedforward motif conductance, especially amongst “superhub” neurons, which promoted the pro-seizure tendency of the network (84). Another computational study using 2P data in the same model estimated microscale avalanche dynamics and showed that there was an increase in network connectivity at single-cell resolution, which drove the brain away from criticality, a point of maximal flexibility in brain state, thus limiting phase transition possibilities until the system converged on an inflexible ictal state (105).

Upon seizure invasion, suprathreshold activity is observed, first as a fast voltage wave and then a slower calcium wave (0.5–1 s delay) (90). Simultaneous calcium imaging through transparent microelectrode and electrocorticography arrays demonstrate spatial concordance between the modalities (106), where progression of ictal electrophysiology is tied with expansion of ictal core (107, 108).

Highly elevated and sustained calcium can be detected (109, 110) in both excitatory cells and inhibitory neurons (95, 102, 111), with pyramidal cells displaying the greatest recruitment and hypersynchrony (80, 89, 103). A concordant expansion of glutamate release into the field is also observed (91, 100), hypothesized to be in part released by glia (98). However, while calcium amplitude is classically thought to be directly correlated with activity, a large increase could also indicate intracellular calcium homeostasis breakdown, and thus an absence of firing. A simultaneous in vivo imaging and patch clamp study revealed that the PV cells with an ictal calcium increase, actually enter a state of depolarization block upon seizure invasion (112), consistent with inhibitory restraint collapse.

Propagation can be witnessed as a calcium wavefront (102, 111, 113), yet the recruitment of interneurons within this seems variable. In the 4-AP model the PV neuron recruitment appears spatially heterogenous (95) at invasion, with a delayed recruitment in SST cells (89). However, in a study imaging Nkx2.1 cells (PV and SST), these were found to be recruited in a spatiotemporally concordant manner to the other cells in the field (114). In a Dravet mouse model with thermally induced seizures, PV and vasoactive intestinal peptide-expressing (VIP) cells appear recruited along with the population, while SST cells demonstrate a spatially independent early recruitment (115). With respect to cortical layers, layer II/III pyramidal cells tend to be the first to propagate during the seizure, with a lagging recruitment of layer V (89, 113). The propagation has also been found to have reliable recruitment across sequential ictal events following intracortical injection of a chemoconvulsant (95, 113). The speed of propagation varied across events, although when the cortex had been first disinhibited by picrotoxin, the speed increased and variability was dampened, consistent with the inhibitory restraint hypothesis (113). Interestingly, in a study where optogenetic photostimulation during calcium imaging was used to determine excitability in individual cells, decreased excitability proximal to the invading seizure wavefront was observed, in contrast to hyperexcitability observed interictally and baseline level excitability observed distal to invasion, suggesting that inhibitory neurons recruited to the seizure generate a front of local inhibition (114).

Propagation through brain regions appears to follow both proximal contiguous and distal homotopic spread with respect to the onset zone (74), including contralateral projections (76, 116, 117). In zebrafish, whole brain propagation was found to typically occur caudal to rostral (78, 100, 118, 119), although the opposite direction was occasionally observed (81), perhaps related to the different developmental stage and chemoconvulsant dosage used (118). While activity in rostral regions (telencephalon and habenula) is not initially correlated with the caudal regions (optic tectum, cerebellum and medulla), entrainment occurs along the rostrocaudal axis upon progression to seizure (98, 99, 120). Indeed, eventually the brain-wide synchrony can be observed (81, 84, 97, 104, 118), although there is some model dependency on the exact extent of recruitment (101, 121). The astroglial syncytium’s calcium activity also displays brain-wide hyperactivity (100) and synchrony within itself and with neurons during seizures (98). There is, though, a short delay in the astroglial ictal recruitment (101), corresponding to a further increase in neural activity, consistent with the hypothesis that neural activity is exacerbated by glial glutamate release (98). Impaired glutamate reuptake by astrocytes also led to hyperexcitability with spontaneous seizures and concurrent excessive glutamate signal (100). At seizure invasion, there is a large increase in intracellular chloride that slowly builds throughout the seizure, while pH slightly decreases (44).

Interestingly, imaging in the visual cortex of a mouse absence model revealed an opposite finding to the other seizure models, a decrease in neuronal activity and synchronization across cortical layers and neuron subtypes during ictal episodes. This asynchronous suppression could be related to the impaired visual awareness classically associated with this seizure type (122).

In the hippocampus, ictal activity shows recruitment in both CA1 pyramidal cells (123, 124), as well as the fasciola cinereum (85). While recruitment is highly synchronous for many neurons at invasion, new neurons are continuously recruited throughout the seizure (125). Spatial propagation dynamics recapitulated those of the cortex, demonstrating within-subject reliability for sequential events and a much faster expansion of electrophysiologic signatures ahead of the neuronal recruitment to the propagating calcium wave (108). Additionally, spatiotemporally concordant release of endocannabinoids occurs, which may play a feedback role in restricting seizure activity (56, 126). Acetylcholine levels have also been shown to increase during seizures, strongly correlating with intracellular calcium (127). Astrocytic calcium increased during seizure invasion as well, sometimes preceding the event, which may be in part mediated by internal store release (86).

With respect to neurovascular coupling, while the ictal focus is typically well supplied and only occasionally hypoxic, the penumbra is hyperoxygenated during seizure initiation and propagation, extending beyond the recruited tissue. At initiation, the blood supply to the focus increases, and this expansion appears as a wave ahead of the calcium wave in neurons into the penumbra and persists into the post-ictal phase (76, 116).

Upon seizure termination or shortly into the post-ictal period, slow propagating waves of calcium have been recorded in the cortex (54, 102, 109, 111, 117) and hippocampus (56, 86, 123, 124). These waves have been hypothesized to be spreading depolarizations as they share similar spatiotemporal propagation properties (128), and spreading depolarizations can be temporally associated with seizures (129, 130). Spatiotemporally concordant serotonin and endocannabinoids waves have also been detected during these calcium waves (54, 56). These calcium waves are then followed by periods of post-ictal suppression of activity and synchrony (118, 124), with one zebrafish study showing functional connectivity to bifurcate into rostral and caudal groups (99). Intracellular chloride returns to baseline levels upon seizure termination and a gradual intracellular acidification occurs post-ictally (44).

A transient increase in astrocytic calcium was also observed post-ictally (86), corresponding to post-ictal vasoconstriction (131), as well as post-ictal hypoactivity (101). While this glial calcium level was sustained for at most a few minutes, vasoconstriction was observed for over an hour. Vasculature smooth muscle cells also showed elevated calcium for the duration of the vasoconstriction (131). When post-ictal vasoconstriction was depressed by a COX-2 inhibitor, the duration of astrocytic calcium elevation was significantly diminished (131), while post-ictal suppression or recovery of neural activity was unchanged (124). On the other hand, when glia glutamate reuptake was impaired, post-ictal hypoactivity was diminished (101). Taken together, a vascular coupling to neuronal and glial activity is present post-ictally, although there may be a dissociation from their post-ictal suppressive mechanisms.

Intravital microscopy coupled with genetically targeted indicators has allowed unprecedented access to the intact nervous system. Leveraging these powerful tools across a variety of vertebrate seizure models has provided deep insight into mechanisms of epilepsy. Imaging has confirmed previous hypotheses derived from electrophysiology, such as the existence of microseizures inside an epileptic focus (132, 133) and the role of inhibition in restraining seizure activity (19, 24, 134). Imaging has also enabled new discoveries which have opened possibilities for novel treatment targets and approaches to epilepsy (17, 135), such as the neuromodulation of adult-born dentate granule cell superhubs (84), or leveraging seizure specific neurochemical changes for drug design (127), which could be extended as an autoregulatory gene therapy (136). Optical and genetic technologies are advancing quickly, opening even more possibilities. Already there are methods that could allow imaging of the whole cortex at a single cell resolution (137, 138), that enable imaging deeper in the brain without the need to aspirate the cortex (139), that capture activity in freely moving subjects to better tie behavior to ensemble activity (119, 140), and to decode electrophysiologic population dynamics from microensemble activity (106, 107).

Collectively, functional optical imaging modalities have immense scalability, to image at the micro-, meso- and macro-circuit level, allowing inference to be drawn about the interplay between the cellular and network evolution of seizures (17, 18, 66). While we think about a classic seizure as evolving, massive hypersynchronous activity, largely due to the use of EEG to identify and study these dynamics, functional optical imaging has made it abundantly clear that while that certainly is a defining feature, there are intricate activity patterns across cell types precipitating and underlying these events.