Optical volumetric brain imaging: speed, depth, and resolution enhancement

When using cameras (or our eyes) for detection, the most straightforward excitation geometry is single-photon wide-field illumination, which was used by Golgi and Cajal when they found branch-like cells in brains. As shown in figure 1(a1), white light illuminates the whole tissue volume along the light path, and transmission or fluorescence signals from the entire sample are captured simultaneously by eyes or cameras. The spatial resolution reaches hundreds of nanometers, which is sufficient to recognize a single neuron or even fiber networks across complex neuron connections. However, wide-field microscopy lacks optical sectioning, and in scattering tissues, suffers from crosstalk, so the imaging depth is typically limited to much below 100 µm in the mouse brain [18]. Below we will discuss different excitation schemes, including single-photon confocal excitation, light-sheet excitation, and multi-photon excitation, that combine with camera-based detection to provide optical sectioning for brain imaging. In addition, a special detection scheme that combines lens array with (mostly) single photon excitation to enable multiple layer imaging simultaneously, i.e. light-field microscopy (LFM), will also be addressed.

2.1.1. (Mostly 1PE) Spinning disk CM.

2.1.2. (Mostly 1PE) LSM.

2.1.3. (Mostly 1PE) LFM.

2.1.4. (2PE) Temporal focusing.

2.1.5. (2PE) Spatially multiplexed two-photon microscopy.

In section 2.1, cameras are combined with wide-field or multi-point excitation schemes to detect excited fluorescence signals. However, wide-field detection is susceptible to signal crosstalk, which limits the imaging depth in scattering tissues. In contrast, methods based on point detectors are less affected by scattering, but with the price of reduced imaging speed due to its scanning nature. It is important to highlight that the data throughput of a single high-speed point detector, such as a PMT or silicon photomultiplier, may achieve ∼1 GHz [78], which is comparable to the total data throughput of the millions of pixels on a camera. Therefore, when combined with beam multiplexing or high-speed scanning, PMT-based methods have the potential to provide high-speed volumetric imaging in tissues. Here we introduce several methods that aim to increase 2D/3D imaging speed by innovative frequency/spatiotemporal multiplexing of excitation beams (section 2.2.1), as well as by high-speed axial scanning techniques (section 2.2.2).

2.2.1. Spatiotemporal multiplexing.

2.2.2. (2PE) High-speed axial scanning.

The most straightforward method for 'deep' tissue imaging is to remove the upper covering tissues. The serial sectioning technique has been applied with EM for a long time, and was introduced into optical imaging a decade ago, under the name of MOST [140]. The concept of MOST is shown in figure 2(a), combining sequentially tissue slicing and imaging, thus allowing submicron voxel resolution for whole mouse brain imaging [140], or 20-micron resolution of a whole human brain [141]. With proper labeling, not only neuron cells, but also vascular distributions down to capillary level could be mapped at 1 µm voxels resolution for the whole mouse barrel cortex [142]. With dual-color imaging based on fluorescence and PI staining, a brain-wide positioning system was established [143].

Zoom In Zoom Out Reset image size

The serial sectioning tomography concept has also been combined with 2PE microscopy [144, 145], to enable whole mouse brain connectome study [146]. In Drosophila, the idea of whole body connectomics was demonstrated, together with novel neural simulation tools [147]. In 2017, block-face serial microscopy (FAST) [148] was developed to significantly enhance the imaging speed, and to enable neuronal circuit mapping in a whole brain of a non-human primate marmoset as well as in a part of a post-mortem human brain.

One major challenge of MOST, and other similar serial sectioning based methods, is to keep connection integrity during slicing, so careful sample preparation design is critical [149]. In the next section, we introduce optical clearing, which can significantly enhance imaging depth of an intact organ or even whole animal without physically cutting the sample.

Tissue clearing techniques improve imaging depth by making the biological specimen transparent via dramatically reducing light scattering, as shown in figure 2(b). The neuroscience community recently witnessed an explosive growth in clearing methods, summarized in an up-to-date review [150]. These methods greatly extend visible light imaging depth toward not only the whole brain, but also the whole animal. Currently, there are three major directions of tissue clearing: solvent-based, aqueous-based, and hydrogel-based. In the following, we briefly describe their clearing capabilities, fluorescence maintenance, volume variation, processing time, and reversibility.

Solvent based clearing methods have been developed for more than 100 years [151], with three major steps including dehydration, lipid removal, and refractive index matching. The techniques are conventionally compatible with chromogenic H&E or metallic Golgi staining, but one major challenge is that dehydration significantly diminishes fluorescence, in particular for fluorescent proteins. Contemporary solvent techniques, such as various versions of 3D imaging of solvent-cleared organs [152–156], enable nearly full transparency in a whole mouse brain, a whole mouse and even a complete human embryo, as well as stabilize the endogenous fluorescent signal. The clearing process typically takes a few hours. It is important to note that solvent-based methods are not reversible, and may shrink the sample size by ∼50%, posing another important challenge to ensure the spatial accuracy of 3D histological structures.

Aqueous-based methods use water-soluble reagents, which form hydrogen bonds with tissue components for refractive index matching, thus providing high levels of biocompatibility, and preservation of fluorescent protein. One representative example is FocusClear, which was demonstrated by Ann-Shyn Chiang nearly 20 year ago [157]. FocusClear can make an insect brain completely transparent within minutes, as well as maintain dye or protein fluorescence with minimal loss. More importantly, the clearing process is reversible, and does not alter the specimen size. The main limitation of FocusClear is its applicability to large samples [158], such as a whole mouse brain. Alternatively, Scale [159, 160] based on urea can clear the whole mouse brain and provides the best fluorescence retention [161]. Nevertheless, tissues in Scale take long incubation time to clear and swell significantly. In general, FocusClear provides better transparency than Scale for small tissues, such as the fly brain.

The third category is hydrogel-based techniques, and a representative example is CLARITY [162], which applies electrophoresis to remove lipids, creating a hydrogel to keep the protein and DNA integrity. It is interesting to note that after electrophoresis, FocusClear is used in CLARITY to make tissue more transparent and maintain fluorescence, thus allowing entire mouse brain imaging via light sheet microscopy. Typically, it takes days to weeks to clear a whole mouse brain, and during electrophoresis some proteins may be lost. Various improvements of CLARITY appeared soon after the first paper, such as passive CLARITY that enabled impressive whole body clearing of a mouse [163], and bone CLARITY that allowed intact bone marrow observation [164]. During the hydrogel processing, the specimen volume may slightly expand, inspiring the idea of expansion microscopy (see section 4.3).

The above clearing methods are applicable only to ex vivo tissues. In vivo clearing methods to reduce the scattering of skin [165, 166] and skull [167] have been developed. In these studies, glycerol is used to match the refractive indices of (collagenous) tissues, so fluorescence is not affected significantly, with no appreciable volume variation. Impressively, the pioneering skull-clearing method, together with two-photon microscopy, offers at least 250 μm imaging depth into the brain pial surface, and is fully reversible. Interested readers are referred to two recent reviews that nicely summarized current state-of-the-art in vivo results [168, 169].

In the above two sections, i.e. MOST and tissue clearing, imaging depth increases via sample processing, so it may alter intrinsic structures. In the following, we address optical techniques that enhance penetration depth and allow in situ and in vivo imaging of brain tissues, including long wavelength modalities, AO, endoscopy, and photoacoustic microscopy.

Imaging depth in a living brain is generally limited due to light attenuation, include scattering and absorption. Since long-wavelength light exhibits less scattering [136, 170], a typical deep imaging strategy is to select long wavelength excitation/emission which avoids water absorbance. For example, by adopting carbon nanotubes whose excitation is at 808 nm and emission at 1300–1400 nm, more than 2 mm imaging depth through an intact skull in a living mouse brain was demonstrated [171].

However, most fluorophores used in brain imaging emit visible light, hence, long-wavelength excitation through multiphoton absorption should be adopted. Multiphoton microscopy provides noninvasive deep-tissue imaging, allows localized excitation in 3D, and exhibits less photobleaching and phototoxicity. The typical imaging depth of a 2PE microscope based on a Ti:sapphire laser is ∼0.6–0.8 mm (section 2.2), much deeper than 1PE in the mouse brain cortex [136]. Therefore, multiphoton techniques have gained a special niche for thick-tissue volumetric imaging.

To further increase imaging depth, several strategies have been developed. One is increasing the Ti:sapphire laser pulse energy, demonstrating ∼1 mm imaging depth in a mouse brain [172]. However, high-energy pulses may induce multiphoton ionization and photodamage. Other strategies adopt long wavelength, high nonlinearity, or the combination. By using a ∼1300 nm excitation source, we have demonstrated third-harmonic generation imaging at 1.5 mm depth in a developing zebrafish embryo [173], and Chris Xu has demonstrated in vivo 2PE imaging up to 1.6 mm depth in a mouse brain [174]. By employing an excitation source at ∼1700 nm, 2PE further pushed imaging depth to 2 mm in vivo [175]. Recently, three-photon microscopy with 1300 nm or 1700 nm excitation wavelength also achieved comparable imaging depth [176, 177], but offers better optical section ability, enabling through-skull and transcutical in vivo imaging in mice and Drosophila, respectively [178, 179]. According to our recent analysis [138], aberration is the main limiting factor of three-photon imaging depth in a Drosophila brain. Therefore, for whole-brain in vivo imaging, three-photon microscopy with AO, which compensates aberrations, shall be the most promising way [180].

Due to inhomogeneous refractive index in biological tissues, the optical wavefronts of both illumination and emission beams are distorted, i.e. optical aberration, resulting in reduced resolution and image contrast. It is well known that the major factor that limits imaging depth is signal-to-background ratio, which degrades in deep tissue due to PSF broadening via scattering and aberration [181]. Scattering may be reduced by long wavelength excitation/emission (section 3.3). To overcome aberration, one possible method is tissue clearing (section 3.2), which may alter the natural sample status. The optical approach to compensate for aberration is through AO, a technique borrowed from astronomical telescopes [182]. Its fundamental working principle is to pre-aberrate the wavefront to compensate for the aberration induced by the specimen/optical system, as shown in figure 2(d). For more details on AO microscopy, we refer the readers to recent comprehensive reviews [183, 184]. Essentially, an AO system is composed of two parts: one for wavefront correction and the other for wavefront characterization, that we briefly introduce below.

For AO wavefront correction in the field of optical microscopy, liquid crystal SLM (LCSLM) [185] or DM [186] are mostly applied. The former offers more complexity in wavefront manipulation, and the latter has higher refresh rates with robust operation independent of polarisation and wavelength. More recently, a plug-and-play multi-actuator adaptive liquid lens is adopted to significantly reduce the setup intricacy and can be applied to commercial microscopes directly [187].

Regarding wavefront characterization, which is necessary to achieve optimal compensation, there are direct and indirect methods. The direct method typically relies on a Shack–Hartmann sensor, composed of a lens array and a camera, to visualize the distorted wavefront from 'guide stars' inside tissues [188]. The indirect one, aka the sensorless method, is based on either iterative optimization of image contrast through Zernike wavefront decomposition [189] or pupil segmentation [190].

In biological tissues like brains, the most commonly encountered aberrations are low-order Zernike modes like spherical aberration, coma, and astigmatism due to index inhomogeneity [191]. Spherical aberrations, as one of the most common aberrations from sample/medium refractive index mismatch, had been extensively studied long before the first adaptive microscope [183]. When various wavefront distortions are effectively corrected by AO systems, image signal-to-background ratio, as well as imaging depth, is improved.

In optically transparent samples, where aberration still exists, AO has shown to recover diffraction-limited images in depth over 270 μm [192]. To extend this approach, near-infrared guide stars are utilized to reduce scattering, enabling high-contrast 2PE imaging down to 700 μm inside the mouse brain [188]. More recently, via microvascular-based near-infrared guide stars, near-diffraction-limited 2PE imaging up to 850 µm in awake mice is achieved, enhancing the signal around 2–3 folds [193]. Furthermore, AO can be integrated in three-photon microscopy to improve image quality up to 1 mm [194].

To increase imaging depth, sections 3.1 and 3.2 consider invasive methods of MOST and clearing, while sections 3.3 and 3.4 address non-invasive methods of long wavelength excitation and AO. In this section, we introduce a partially invasive method, optical endoscopy, which may achieve centimeter imaging depth, while maintaining subcellular spatial resolution. Endoscopy observes the inner part of an organ via inserting thin imaging probes, including multicore fiber bundles, a single-mode fiber, a double-cladding fiber, a multi-mode fiber (MMF), or a gradient-index (GRIN) lens [195, 196]. Here we mainly focus on multiphoton endoscopy, where the GRIN lens and MMF were adopted in most cases for volumetric in vivo imaging of the brain, as shown in figure 2(e).

GRIN-lens-based multiphoton endoscopy was first demonstrated in 2003, with a GRIN lens of 1–2 cm length, i.e. the reachable imaging depth, in a zebrafish brain [197], and later in a mouse brain [198, 199]. An amazing case with unprecedented imaging depth uses a 28.5 cm long GRIN lens for two-photon imaging of unstained tissue, manifesting that multiphoton endoscopy for a human brain might be possible [200]. The 3D spatial resolution of multiphoton endoscopy can be improved when the intrinsic aberration of the GRIN lens is corrected [201].

More recently, GRIN lens is combined with various multiphoton volumetric approaches that we introduced in section 2. For example, Moretti et al used phase modulation to efficiently project complex 2PE patterns in the focal plane of a GRIN lens, performing scanless functional volumetric imaging (cf section 2.1.5) at ∼1.2 mm depth in rodent hippocampal networks in vivo [202]. Sato et al presented a fast varifocal two-photon microendoscope system equipped with a GRIN lens and an ETL (cf section 2.2.2) [203], enabling quasi-simultaneous neuronal activity mapping of two deep-brain focal planes separated by 85–120 µm at ∼10 fps. In 2019, Meng et al incorporated a Bessel focus scanning module (cf section 2.2.2) into two-photon fluorescence microendoscopy, with lateral resolution improved, to observe mouse hippocampus and lateral hypothalamus in vivo [204]. In the open source 'Miniscope' project, the latest development is to optimize light-field acquisition (cf section 2.1.3) to achieve 40 fps 3D brain imaging in freely-moving animals [205]. In 2021, we combined TAG lens (section 2.2.2) and a GRIN lens to observe calcium dynamics in suprachiasmatic nuclei, which is located at the bottom of a mouse brain [206].

Apart from the GRIN lens with nearly mm diameter, a hair-thin MMF with ∼100 μm diameter serves as a much less invasive endoscopic device. In the past, MMF was mainly used to deliver light instead of image acquisition. Thanks to recent advances in wavefront shaping that allows proper focusing after an MMF, MMF-based multiphoton endoscopy was realized [207–209], and fast volumetric imaging has also been achieved [210, 211]. The main drawback of MMF endoscopy is its limited FOV, which may improve via novel AO approaches [212].

The dream of in vivo brain imaging is to reach the whole brain depth, say centimeter scale, and maintain cellular spatial resolution and tissues intact, i.e. without the need of skull and scalp removal. Photoacoustic imaging is a potential candidate toward this goal. The essence of photoacoustic imaging is a combination of optical excitation and acoustic detection, as illustrated in figure 2(f), which allows in vivo label-free imaging of the brain [213, 214]. In photoacoustic imaging, biological tissues are excited using broad or focused laser beams with appropriate wavelengths to the targeted optical absorbers, e.g. hemoglobin. The optical absorbers absorb and then convert the laser energy into ultrasound waves via thermoelastic expansion. The photoacoustic image with high optical absorption contrast is then formed with the detected ultrasound. Leveraging optical absorption contrast and relatively weak acoustic attenuation and scattering (compared with optical ones) in biological tissues, photoacoustic imaging has been demonstrated to be capable of performing real time morphological, functional and molecular imaging of living subjects with high contrast and centimeter-scale penetration [215, 216].

Since Wang et al reported the first photoacoustic imaging work on in vivo noninvasive transdermal and transcranial imaging of the structure and function of rat brains, photoacoustic imaging has gained growing attention in brain imaging [217]. Photoacoustic brain imaging nicely complements the drawbacks of other deep brain imaging modalities, e.g. poor depth-to-resolution ratio in diffusive optical tomography and low temporal resolution in MRI. The spatial resolution, imaging depth, and functional capability of the photoacoustic brain imaging mainly depends on the configuration of laser excitation and acoustic detection.

With excitation of broad laser beams, two types of photoacoustic imaging—photoacoustic tomography (PAT) and acoustic resolution photoacoustic microscopy (ARPAM) are proposed. In PAT and ARPAM, because the absorbed photons are localized ultrasonically, the en face and axial spatial resolution is limited by acoustic diffraction and detector bandwidth, respectively, ranging from few tens to few hundreds of micrometers. Generally, PAT is equipped with low frequency ultrasound transducers (e.g. <10 MHz) and can offer > 5 cm penetration because low frequency ultrasound suffers less attenuation [216]; ARPAM using high frequency transducers (e.g. >20 MHz) can provide 1 cm penetration. Though there is trade-off between the penetration and spatial resolution in PAT and ARPAM, depending on the detected ultrasound frequency, the depth-over-resolution ratio of both PAT and ARPAM is larger than 100, indicating that PAT and ARPAM are still high resolution imaging techniques [213]. The uniqueness is that with endogenous hemoglobin contrast, PAT and ARPAM can image dynamic changes of multi-neurovascular-coupling parameters independently, such as total hemoglobin concentration, cerebral blood volume, and cerebral blood oxygenation, in single blood vessel level over the whole rat and mouse brain without the need of skull removal [218, 219]. With endogenous contrast, e.g. genetically-encoded probes plus calcium dependent absorption spectra, PAT can offer deep molecular imaging of neural activity [220, 221].

On the other hand, with focused-beam excitation, optical resolution photoacoustic microscopy (ORPAM) has been developed. With optical focusing, the en face spatial resolution of ORPAM corresponds to the optical diffraction limit, reaching ∼1 μm and even sub-micrometer while the imaging depth is compromised to within one-transport mean free path, i.e. ∼1 mm. Moreover, because of the acoustic detection nature, the axial resolution of ORPAM is still limited by acoustic detection bandwidth, which is about few tens of micrometers when a 50 MHz high frequency transducer is used [213]. Combining blood flow information with the multi-neurovascular-coupling parameters mentioned above, it has been demonstrated that ORPAM is capable of transcranial cortical imaging of cerebral metabolic rate of oxygen in single micro-vessel level in mouse brains with endogenous hemoglobin contrast [222]. ORPAM also has shown its capability of trans-cuticle brain imaging of Drosophila with exogenous fluorescence protein because of the low acoustic attenuation of the cuticle [223].

With centimeter-scale penetration and high spatial resolution, moving toward PAT human brain imaging is the most exciting and challenging research direction [224]. Although the optical and acoustic attenuation of the thick human skull significantly degrades the PAT human brain imaging quality, significant progress is expected in coming years [225].

To conclude section 3, we summarized the performance of depth-enhancing techniques in table 2. Starting from techniques that are mostly applied to ex vivo brain tissues, i.e. MOST and tissue clearing, the imaging depth of the former is in principle unlimited, and the latter can achieve at least whole mouse body clearing. MOST maintains the spatial resolution throughout the sections, and may achieve depth-over-resolution ratio above 10000. When combined with super-resolution modality, the ratio exceeds 100 000! However, for large samples, it is time consuming and faces the difficulty of maintaining sophisticated connections. Optical clearing may preserve the neural patterns, whereas in a large tissue, the spatial resolution is degraded with depth. Combining the advantage of these two methods would enable whole brain/animal neural structure mapping with best optical resolution and intact fiber structures.

For in vivo methods, long-wavelength excitation in the biological transparent window provides 1–2 mm penetration depth together with <1 μm spatial resolution, achieving depth-over-resolution ratio above 1000. Combining with AO instruments, the image contrast, as well as spatial resolution, can be further enhanced. Although the contrast enhancement should lead to significant depth enhancement, currently this potential has not been fully utilized, and imaging depth of noninvasive optical microscopy is limited to mm-scale. To further enhance depth, a different contrast based on photoacoustic detection may be adopted, providing multi-scale imaging capability, from whole-mouse imaging with 100 μm resolution, to 1 mm depth with 1 μm resolution (similar to pure optical imaging).

Overall speaking, for noninvasive optical imaging modalities in table 2, resolution degrades as imaging depth increases, leading to depth-over-resolution ratios of a few hundreds to one thousand. This trade-off between depth and resolution is well documented in literature [226, 227]. As shown in table 2, invasive methods, such as MOST and clearing, break the trade-off, but sacrifice the in vivo observation ability. The partially invasive method, endoscopy, breaks the trade-off by offering penetration depths larger than 1 cm and spatial resolution around 1 μm, while maintaining in vivo imaging capability. In the future, innovative in vivo clearing methods may bring breakthroughs in this field. Another potential candidate to achieve in vivo observation while breaking the resolution-depth trade-off is artificial intelligence. Through a training set consisting of volumetric blurred and high-resolution serial images, deep-tissue resolution may be considerably enhanced [228]. This is currently feasible only for ex vivo imaging. Applying deep-learning to in vivo imaging may lead to artifacts due to previously unknown patterns.

It is important to note that the depth-resolution trade-off exists not only within optical imaging, but is a general phenomenon across different imaging modalities. For example, medical imaging modalities such as MRI, computed tomography, positron emission tomography, and ultrasound, have human-body penetration capability, but their resolutions are generally at submillimeter to millimeter scale. Optical imaging modalities such as optical coherence tomography and diffuse optical tomography offer 1–10 cm depth, but the resolutions are not adequate to resolve cells. EM has nanometer resolution, yet it is limited to surface or ultrathin slice imaging. By further studying and comparing physical origins of depth-resolution limitation among different techniques, new ideas of optical imaging might emerge. For example, the resolution of MRI is determined by magnetic field gradient, not affected by tissue scattering, and reaches two orders smaller than its radio-frequency electromagnetic wavelength. This may be an inspiring example for optical imaging to break the diffraction limit in scattering tissues, which is the topic of next section.

The deterministic method features a patterned excitation to selectively turn 'on' the molecules. Notable examples of deterministic methods include stimulated emission depletion (STED) [233], reversible saturable optical fluorescence transitions (RESOLFT) [234], excitation nonlinearity [235], and structured illumination microscopy (SIM) [236], that we are going to discuss respectively below.

4.1.1. STED/RESOLFT.

A characteristic of STED and RESOLFT microscopies (figure 3(a)) is the need of two illumination beams: one for excitation and the other for inhibition. The inhibition beam, which is donut-shaped, turns 'off' peripheral fluorescence emission and confines the excitation of fluorophores ('on' state) in a small volume, achieving ∼20 nm resolution in live cell imaging [237–239]. Three dimensional super-resolution microscopies are also developed for live brain imaging and its resolution is ∼60 nm laterally and ∼150 nm axially [240, 241]. Similar resolution enhancement via confinement of emission can also be achieved by non-fluorescent nanomaterials, such as absorption of graphene [242] and scattering of gold nanoparticles [243].

Zoom In Zoom Out Reset image size

For tissue imaging, aberration and scattering distort the donut inhibition beam, thus deteriorating super-resolution. As mentioned in section 3, imaging depth can be significantly enhanced by reducing aberration and scattering. By correcting spherical aberration with objective correction collar, STED achieves 70 nm resolution at 120 μm depth in organotypic brain slices [244]. Through reducing scattering via long-wavelength excitation, two-photon STED demonstrated 60 nm resolution at 50 μm depth in acute brain slices [245, 246]. Combining STED with a hollow Bessel beam, which is less susceptible to scattering, provides >100 μm imaging depth in a phantom. Although the same strategies should all be applicable to RESOLFT [247], the reported imaging depth of RESOLFT to date is only about 30 μm [248].

4.1.2. Excitation nonlinearity.

4.1.3. Structured illumination microscopy (SIM).

When adjacent fluorophores emit simultaneously, as in conventional microscopy, they are not resolvable. However, if they emit 'stochastically', leveraging fluorescence information accumulated within hundreds of frames, allows to localize the spatial coordinate with a precision up to 10 nm. The resolution enhancement is obtained through fitting the center of emission PSF (sparsely). This is the underlying principle of stochastic super-resolution imaging techniques, presented in figure 3(d). There are many different variants, just to name a few well-known examples like stochastic reconstruction microscopy (STORM) [262], photoactivated localization microscopy (PALM) [263], and DNA point accumulation for imaging in nanoscale topography (DNA-PAINT) [264]. They are all based on wide-field detection schemes, so to achieve tissue imaging, it is necessary to incorporate optical sectioning techniques (cf section 2.1), such as LSM and LLSM in figure 3(d1) [265, 266]. Adding long-wavelength excitation and two-photon excitation with LSM plus PALM, 40 nm resolution is achieved in a 150 μm thick agarose scattering phantom [267]. Stochastic super-resolution has also been combined with single-objective LSM with oblique illumination, and demonstrated 40 nm resolution at 40 μm depth in mouse brain slices [268].

In principle, all the camera-based volumetric imaging techniques in section 2.1 should be compatible with stochastic super-resolution. For example, temporal focusing has been combined with PALM (figure 3(d2) [269]); SDCM has been combined with DNA-PAINT (figure 3(d3) [270]) to significantly improve the spatial resolution. However, these reports used cell samples, so the demonstrated imaging depth is below 10 µm. An innovative approach based on fluorescence self-interference (SELFI) to capture the emitter's axial position allows impressively 40 nm isotropic resolution at 50 μm depth of stem cell organoids [271].

As mentioned in section 3.2, one strategy to significantly enhance penetration depth is through optical clearing. A notable example is light-sheet localization microscopy for clarified tissue (LLM-CT), where the whole Drosophila brain can be mapped with 50 nm resolution within 1 day [272]. This is orders-of-magnitude faster than EM [4]. Confocal localization deep imaging with Optical clearing (COOL), a technique combining SDCM and tissue clearing, can achieve 20 nm resolution (localization uncertainty) across the whole Drosophila brain (∼200 µm) [273], i.e. achieving depth-over-resolution ratio of 10 000.

Expansion microscopy, differently from other techniques which exploit the illumination or detection modification, enhances spatial resolution via isotropic physical expansion of biological tissues [274], schematically shown in figure 3(f). The technique is a combination of hydrogel clearing (cf section 3.2) and the physics of swellable polyelectrolyte hydrogels. The former allows preservation of tissue staining, while the latter offers significant resolution enhancement (typically 4× to 10×) with conventional microscopes [275]. With 10× expansion, the attainable resolution is about 20 nm, which is comparable to other super-resolution techniques. However, the limitation arises from the working distance and numerical aperture (NA) trade-off, which is similar to the general trade-off between penetration depth and resolution that we discussed in the last two paragraphs of section 3. When a high NA objective is adopted to enhance the spatial resolution, the working distance is often short. For example, a 40 × 1.15 NA water-immersion objective has 600 µm working distance. When using this objective with a 20-fold iterative expanded sample, the effective imaging depth is merely 30 µm before expansion [276]. By contrast, a four-fold magnified analysis of the proteome, a similar version of expansion microscopy without protease treatment, adopts a low NA (0.6) long working distance (8 mm) objective but the resolution degrades quickly in the tissues deeper than 100 µm due to the inherent aberration [277]. Currently, the-state-of-the-art expansion microscopy for super-resolution imaging works only for small-animal brains, such as a whole Drosophila brain with 60 × 60 × 90 nm resolution at 4× expansion [35]. For large samples, either sectioning should be conducted before imaging (section 3.1) or imaging at low magnification under the new Mesolens with long-working distance and high NA is utilized [278].

This review has covered imaging speed in section 2, depth in section 3, and resolution enhancement in section 4. In sections 4.1–4.3, we reviewed recent advances of super-resolution imaging in thick samples, i.e. addressing the factors of resolution and imaging depth. Now in this section, we further consider the factor of imaging speed, toward the vision of high-speed super-resolution volumetric imaging in brain.

For deterministic approaches, which mostly rely on point scanning, the techniques that we addressed in section 2.2, including multiplexing and high-speed scanning, shall be applicable to improve the speed. The first spatially multiplexed STED and RESOLFT were both developed by Stefan Hell's group [279, 280]. High-speed scanners such as an electro-optical scanning (>1000 frame s⁻¹) was combined with STED to enable high-speed tracking of vesicles with <10 ms temporal resolution [281]. A very recent work combining electro-optical scanning and multiplexing shows promising millisecond STED imaging over large fields of view [282]. On the other hand, the axial resolution (but not lateral resolution) of high-speed light sheet microscopy can be enhanced by one order of magnitude via combination with RESOLFT [283] and STED [284].

For stochastic super-resolution methods, the speed limitation is due to the requirement of accumulating hundreds of frames. Artificial neural networks, by requiring its powerful image recognition capability [285], accelerate super-resolved image reconstruction with two-order less raw images [286] or with faster computation [287]. With accumulation of large pre-training dataset, even a single diffraction-limited image can be transformed into a super-resolution image after a deep learning neural network [288]. Nevertheless, it may be difficult or time-consuming to prepare a large training dataset.

For expansion microscopy, in principle, all the high-speed modalities mentioned in section 2 can be applied. As we mentioned earlier, the combination of expansion microscopy and lattice LSM (ExLLSM) [35] provides synapse-level resolution over an entire Drosophila brain within a few days. It is interesting to notice that LLM-CT (section 4.2) is superior both in terms of spatial resolution and speed. We envision that the combination of stochastic super-resolution microscopy, light sheet illumination, and expansion sample processing may provide ultimate performance in resolution, speed, and imaging depth, pushing toward the circuit visualization of not only mammalian brains, but even a complete human brain, down to synaptic level. Note that serial sectioning may be required for such large samples, and currently expansion (or optical clearing) is not compatible with living brains.

To conclude this super-resolution section, a comparison of aforementioned super-resolution techniques is presented in table 3 and figure 4. Although the performance (e.g. resolution, depth, etc) of each technique is sample-dependent, it is interesting to note that in general, deterministic techniques (STED, nonlinearity, and SIM) provide faster speed, i.e. less than 1 s acquisition time per FOV, and less spatial resolution of 60 nm or above; stochastic methods (STORM, PALM, DNA-PAINT, and SELFI) have superior resolution around 20 nm, but take more than 100 s to accumulate enough frames for a single reconstructed FOV, implying a trade-off between speed and resolution. Expansion microscopy is between these two categories, nevertheless its sample preparation may be more time-consuming.

Zoom In Zoom Out Reset image size

Another common trade-off in the field of optical microscopy is between depth and resolution, as mentioned in the conclusion of section 3 (table 2). Here in table 3, the ratio between depth and xy resolution is presented for quantitative comparison of different super-resolution imaging modalities. These ratios of most techniques are on the order of 1000, similar to that of noninvasive techniques in table 2, manifesting the ubiquitous depth-over-resolution trade-off. In table 3, LLM-CT and COOL rely on tissue clearing (section 3.2), leading to one order enhancement of depth-over-resolution performance. Interestingly, table 2 also displays a similar depth-over-resolution enhancement via clearing. When considering 3D resolution, SELFI features the smallest 3D PSF, as well as highest depth-over-PSF-volume ratio, and more importantly, it is feasible for in vivo imaging, though the speed is slow. More technical innovations are highly desirable toward the goal of in vivo whole brain imaging at 10–20 nm spatial resolution.