Brain activity can be quantified either directly, through measuring the generation of action potentials, or indirectly, through measuring changes in cytosolic protein products or changes in calcium level post neuron firing. Transcription of c-fos in neurons (Curran and Teich, 1982) is rapidly increased when neurons are activated (Krukoff, 1999). This allows c-fos to be used as an indicator of neural activity in response to stimulation, and thus the detection of brain regions that are involved in specific neural processing tasks in zebrafish (Decarvalho et al., 2013; Kappel et al., 2022). This can be done by measuring either c-fos protein translation via immunohistochemistry or mRNA expression using in situ hybridization.

Intracellular calcium levels can also be used as a proxy for neural activity. During an action potential the opening of voltage-gated calcium channels causes a 10-fold increase in free calcium within 1 ms (Luo et al., 2018). The calcium influx then leads to phosphorylation of mitogen-activated protein kinase [MAP, also known as pERK, (Rosen et al., 1994)]. Immunohistochemical detection of pERK can therefore be used as a measure for neural activity (Randlett et al., 2015; Corradi et al., 2022). Alternatively, sensor protein-based detection of calcium can also be used which enables live visualization of neural activity (Whitaker, 2010). The most commonly used calcium indicators are the GCaMP family (Grienberger and Konnerth, 2012). These proteins consist of three major components, the circularly permuted green fluoresent protein (GFP), the calcium-binding protein calmodulin, and a calcium-bounded-calmodulin-interacting M13 peptide (Nakai et al., 2001; Akerboom et al., 2012). The binding of calcium to calmodulin causes a conformational change in the GCaMP and releases the GFP from the protonated state, increasing the fluorescence level. Several generations of improvement have been made to increase the sensitivity, speed and signal of the GCaMPs (Nakai et al., 2001; Tian et al., 2009; Chen et al., 2013). Zebrafish can avoid some of the challenges of GCaMPs in rodent models, such as uneven expression (Sakamoto et al., 2022). In zebrafish pan-neuronal expression can be achieved via neuroD or HuC promotors, which are critical gene expressed during neurogenesis (Park et al., 2000; Rupprecht et al., 2016; Oldfield et al., 2020). GCaMPs can be targeted to the nucleus by fusing them to human histone H2B protein (Kanda et al., 1998; Vladimirov et al., 2014). In zebrafish neurons the nucleus occupies almost the entire soma, making this a very effective technique as it facilitates image segmentation compared to targeting GCaMPs to the cytosol. GCaMPs have helped reveal large-scale population response patterns in zebrafish (Del Bene et al., 2010; Muto et al., 2011; Vladimirov et al., 2014). Ratiometric indicators can also be used for visualizing neural activity via fluorescence resonance energy transfer (FRET). Ratiometric indicators such as cameleon consist of two fluorophores. Binding of calcium activates the emission from one of these via FRET, resulting in changes in the fluorescence ratio between the two fluorophores, which can then be used as a proxy of neural activity (Mank and Griesbeck, 2008; Kettunen, 2012). Ratiometric indicators are particularly useful for measuring basal-level activities and have been used in functional mapping and connectivity mapping studies in zebrafish (Li et al., 2005; Tao et al., 2011).

Although GCaMPs are very powerful they are relatively slow, due to the intrinsic dynamics of calcium in neurons and the dynamics of the indicators themselves (Ali and Kwan, 2020). A more direct measure of neural activity is to visualize changes in the membrane potential via genetically-encoded voltage indicators (Xu et al., 2017; Luo et al., 2018; Abdelfattah et al., 2019), which come in two main types. The first type utilizes the voltage-sensing domains of voltage-sensitive phosphatase or ion channels in the membrane. By fusing GFPs with these voltage-sensing domains, the voltage-induced conformational change results in changes in fluorescence (St-Pierre et al., 2014). The second type takes advantage of microbial rhodopsins which produce fluorescence responses to voltages changes (Kralj et al., 2012). Voltage indicators can resolve single action potentials (Kralj et al., 2012; Luo et al., 2018), and combining voltage imaging with cell-type specific markers has helped reveal fine spatial and temporal details of neural activity in zebrafish such as the propagation of motor signals in spinal cord circuits (Böhm et al., 2022). However signals from voltage imaging are generally weaker than those from calcium imaging, and the signal-detection challenge is increased by the high imaging rates necessary to capture the fast temporal dynamics of membrane-voltage fluctuations.

Visualization of neural activity at the synaptic level can be achieved via genetically-encoded neurotransmitter sensors (Wang et al., 2018; Vogt, 2019). Sensing of neurotransmitter release can be achieved either via bacterial-derived binding proteins or via G-protein-coupled receptors (GPCRs) with specific selectivity (Marvin et al., 2013; Jing et al., 2018; Marvin et al., 2018; Patriarchi et al., 2018; Sun et al., 2018; Marvin et al., 2019). Visualization of neurotransmitter activity can be achieved by engineering an insertion of fluorescence protein on to these sensor proteins.

Several systems have been developed to allow cell-type specific expression of fluorescence labels including the Gal4 (a yeast transcription factor) and Gal4 upstream activating sequence (gUAS) system (Gal4/gUAS), the tryptophan repressor (utilizing tryptophan biogenesis in E. coli) and its UAS system (TrpR/tUAS), and the QF transcription factor (adopted from fungus Neurospora crassa) and its UAS system (QF/QUAS). All these systems were developed to create a binary expression system where the fluorescence reporter protein expression is regulated by cell-type specific promoters (Asakawa and Kawakami, 2008; Suli et al., 2014; Ghosh and Halpern, 2016). Limitations of these systems include DNA methylation over generations which leads to gene silencing in some lines of Gal4/gUAS system, and toxicity when expressed by strong promoters in the TrpR/tUAS system (Suli et al., 2014; Burgess et al., 2020). Nevertheless, these systems have been used widely in zebrafish for targeted expression with calcium indicators (Choi et al., 2021; Barker et al., 2021), voltage indicators (Böhm et al., 2022) and neurotransmitter indicators (Yoshimatsu et al., 2020).

nacre zebrafish (which carry mutations that affect pigment cells) are transparent at larval stages, making them ideal for volumetric optical imaging of brain activity (Huang et al., 2021). Casper (White et al., 2008) and Crystal (Antinucci and Hindges, 2016) fish maintain this transparency in adulthood. Since up to about 15 dpf zebrafish breathe through their skin, before this age the unparalyzed and unanaesthetized larvae can be easily immobilized in agarose for imaging. Although zebrafish start to hunt live prey at around 5 dpf, they can survive without any external source of food for several days beyond this (Hernandez et al., 2018; Lagogiannis et al., 2020); thus imaging can occur for many hours or even days at a time with no maintenance of the animal required. To allow some degree of behavioral output, the tail can be freed while maintaining the head immobile under the microscope objective. Virtual reality environments can be created and tail movements (or electrode recordings of motorneuron activity) used to drive movement through these environments (Ahrens et al., 2012; Trivedi and Bollmann, 2013; Vladimirov et al., 2014; Torigoe et al., 2021). Alternatively stimuli can be projected onto a screen in front, to the side or beneath the fish. Looming stimuli and also small, moving or stationary prey-like spots can evoke strong neural responses in several brain areas, most notably the optic tectum, and also tail movements (Semmelhack et al., 2014; Bianco and Engert, 2015; Fernandes et al., 2021). In addition, some studies have used head-fixed fish observing free-swimming paramecia (Muto and Kawakami, 2013; Wee et al., 2019; Oldfield et al., 2020).

The principal techniques that have been applied to brain imaging in head-fixed fish are spinning disk, 2-photon, light sheet, and 2-photon light sheet microscopy. The main challenge is to image fast enough so as to capture volumes of all or at least a substantial fraction of the brain at rates exceeding 1 volume/s. The larval zebrafish brain spans dimensions of roughly 450 μm width by 700 μm length by 320 μm height (Svara et al., 2022). “Whole brain” imaging in this context generally means the full extent of x and y but more limited z, which is sufficient to provide complete imaging of most brain regions except the deepest. Sophisticated automated techniques are then required to convert the raw imaging data to df/f activity traces such as CaImAn (Giovannucci et al., 2019) and Suite2P (Pachitariu et al., 2017).

Two main LFM modalities are known as conventional and Fourier LFM. In conventional LFM the MLA is placed in the native image plane and the camera is placed behind the MLA at a distance of one microlens focal length (Figures 3A, B). Here each pixel captures a unique ray, i.e., view, emitted at a specific angle within the numerical aperture of the microscope objective. Hence, the angular resolution is equal to the number of pixels behind each microlens. The lateral resolution is equal to the microlens pitch divided by the objective magnification (Broxton et al., 2013). This design is computationally expensive due to the use of spatially varying point spread functions (PSFs). Placing the MLA in the imaging plane also produces more reconstruction artifacts near the center of the imaging volume leading to loss of resolution in the center.

Conventional LFM has only recently been applied within neuroscience, and has enabled simultaneous functional imaging of neuronal activity at single-cell resolution in an entire Caenorhabditis elegans nervous system, the whole larval zebrafish brain (Prevedel et al., 2014; Andalman et al., 2019), and large brain volumes in mice in vivo during behavior (Grosenick et al., 2017). In immobilized zebrafish, conventional LFM has achieved a spatial resolution of 3.6 × 3.6 × 5.0 μm³ across a reconstruction volume of 700 × 800 × 500 μm³ at a volume imaging rate of 5 Hz (Andalman et al., 2019).

In Fourier LFM the MLA is placed in the Fourier plane, i.e. the back aperture of the microscope objective, rather than the imaging plane itself (Figure 3C). This is done by transforming the image formed at the native image plane to the back focal plane using a Fourier lens, with the camera at the back focal plane of the MLA (Guo et al., 2019). This design allows a 2D spatially-invariant PSF to be defined, which facilitates simpler and more efficient deconvolution algorithms. It also reduces the reconstruction artifacts near the imaging plane as the MLA is placed away from the imaging plane.

Cong et al. (2017) extended the FoV further by using a customized MLA, which they termed eXtended LFM or XLFM (Figure 3D). The customized MLA was divided into 2 groups with an axial displacement between them to extend the range of depth. Using this technique Cong et al. (2017) achieved a maximum resolution of 3.4 μm by 3.4 μm by 5 μm and reconstructed a lateral extent of 800 μm and an axial extent of 400 μm in freely-moving zebrafish. Fluorescence was excited using 200 μs pulses from an LED, and the system was capable of a volume imaging rate of 77 Hz. Using XLFM as the hardware, Yoon et al. (2020) further improved the XLFM reconstruction performance by increasing the SNR from 9 to 74 at volume imaging rates of up to 50Hz.

Unlike most scanning-based volumetric microscopy methods, where the volumetric reconstruction process is essentially a stacking operation of the deconvoled 2D slice images, LFM systems rely heavily on computational processing to recover volume information. Since voxels at different depths overlap on the camera sensor, 3D deconvolution algorithms are required to infer volumetric information from the light field measurement. The goal of a LFM deconvolution algorithm is to estimate the volumetric light radiant distribution ĝ of the target sample g that was most likely to have generated the recorded light field image f on the camera sensor. Assuming a linear image model H, the deconvolution can be expressed concisely as a matrix inversion problem:

Here H can be composed by measuring or simulating the PSFs of the system. If the PSFs are shift-invariant at the same axial depth, the matrix multiplication operation can be reduced to a convolution operation. This assumption holds for Fourier LFM systems due to the consistently-aliased spatial and angular information in the Fourier domain (Guo et al., 2019). Deconvolution algorithms can be divided into model-driven and data-driven methods depending on whether the deconvolution process requires a training dataset.

Model-driven methods focus on modeling the physics of image formation instead of relying on a-priori knowledge of the light field image dataset f distribution. Here the forward model H is the PSF measurement matrix, and the deconvolution equation above represents a non-blind deconvolution problem. To store and apply the large H matrix, the repeating patterns and the sparsity of the PSFs structure are usually exploited to reduce the cost of the inversion problem.

The Richardson-Lucy (RL) deconvolution algorithm applied to LFM reconstructs the 3D volume by maximizing the Poisson likelihood of the light-field image given an estimate of the 3D volume (Richardson, 1972; Lucy, 1974). The RL deconvolution algorithm and its variants take the PSF stack and LFM image as input and are thus non-blind. The algorithm relies on gradient descent to estimate the most likely 3D volume that produced an observed LFM image (Figure 4A). Due to their robustness and wide implementation support, they are the most popular methods for LFM applied to zebrafish brain imaging (Broxton et al., 2013; Cohen et al., 2014; Prevedel et al., 2014; Perez et al., 2015; Cong et al., 2017; Nöbauer et al., 2017; Taylor et al., 2018; Truong et al., 2020; Zhang et al., 2021a). However, since the iteration process is time-consuming, RL methods are not useful for real-time and high-throughput LFM use cases.

The sparsity of neuronal spiking can be exploited to increase deconvolution performance. For instance, Pégard et al. (2016) and Yoon et al. (2020) demonstrated enhanced resolution by minimizing the error between simulated LFM images of sparse neuronal spiking and observed LFM images using an accelerated proximal gradient algorithm. In Yoon et al. (2020), the authors first extracted the sparse neuron positions with the alternating direction method of multipliers (Candes et al., 2009) and then used a forward model to simulate LFM images. By applying Richardson-Lucy deconvolution on simulated sparse LFM images instead of observed LFM images, the final resolution and deconvolution time was improved.

Data-driven methods infer the mapping relationship between the 3D sample volume and an LFM image from a dataset of image-volume pairs. This is usually done by utilizing deep neural networks (DNNs) due to their powerful non-linear fitting capacity. The PSF measurement information here is optional as long as the methods can implicitly learn the underlying physical process H⁻¹ correctly. However DNNs require a large amount of training data, rely on numerous parameters, and usually require long training times. However, in return they offer higher reconstruction accuracy and real-time inference speed (Wagner et al., 2021). Wang Z. et al. (2021) first applied a variant U-Net DNN (Ronneberger et al., 2015) to a blind LFM deconvolution algorithm and achieved real-time zebrafish brain volumetric reconstruction (Figure 4B). Wagner et al. (2021) then introduced a hybrid light-sheet and LFM system they termed HyLFM, which used light-sheet microscopy data to supervise a DNN-based blind deconvolution algorithm for LFM images. This method improved blind deconvolution performance because both LFM and light-sheet data were captured with the same optics, and therefore implicitly used the same measurement matrix at the time of capture. While model-based LFM deconvolution algorithms are still in their infancy, future improvements will likely incorporate ideas from newly-emerging physics-informed machine learning techniques (Zhang and Ghanem, 2018; Shlezinger et al., 2019; Ongie et al., 2020; Khan et al., 2022; Yanny et al., 2022).