Model-Based Explainable Deep Learning for Light-Field Microscopy Imaging.

In modern neuroscience, observing the dynamics of large populations of neurons is a critical step of understanding how networks of neurons process information. Light-field microscopy (LFM) has emerged as a type of scanless, high-speed, three-dimensional (3D) imaging tool, particularly attractive for this purpose. Imaging neuronal activity using LFM calls for the development of novel computational approaches that fully exploit domain knowledge embedded in physics and optics models, as well as enabling high interpretability and transparency. To this end, we propose a model-based explainable deep learning approach for LFM. Different from purely data-driven methods, the proposed approach integrates wave-optics theory, sparse representation and non-linear optimization with the artificial neural network. In particular, the architecture of the proposed neural network is designed following precise signal and optimization models. Moreover, the network's parameters are learned from a training dataset using a novel training strategy that integrates layer-wise training with tailored knowledge distillation. Such design allows the network to take advantage of domain knowledge and learned new features. It combines the benefit of both model-based and learning-based methods, thereby contributing to superior interpretability, transparency and performance. By evaluating on both structural and functional LFM data obtained from scattering mammalian brain tissues, we demonstrate the capabilities of the proposed approach to achieve fast, robust 3D localization of neuron sources and accurate neural activity identification.

Voltage imaging reveals the dynamic electrical signatures of human breast cancer cells.

Cancer cells feature a resting membrane potential (Vm) that is depolarized compared to normal cells, and express active ionic conductances, which factor directly in their pathophysiological behavior. Despite similarities to 'excitable' tissues, relatively little is known about cancer cell Vm dynamics. Here high-throughput, cellular-resolution Vm imaging reveals that Vm fluctuates dynamically in several breast cancer cell lines compared to non-cancerous MCF-10A cells. We characterize Vm fluctuations of hundreds of human triple-negative breast cancer MDA-MB-231 cells. By quantifying their Dynamic Electrical Signatures (DESs) through an unsupervised machine-learning protocol, we identify four classes ranging from "noisy" to "blinking/waving". The Vm of MDA-MB-231 cells exhibits spontaneous, transient hyperpolarizations inhibited by the voltage-gated sodium channel blocker tetrodotoxin, and by calcium-activated potassium channel inhibitors apamin and iberiotoxin. The Vm of MCF-10A cells is comparatively static, but fluctuations increase following treatment with transforming growth factor-ß1, a canonical inducer of the epithelial-to-mesenchymal transition. These data suggest that the ability to generate Vm fluctuations may be a property of hybrid epithelial-mesenchymal cells or those originated from luminal progenitors.

Comparing synthetic refocusing to deconvolution for the extraction of neuronal calcium transients from light fields.

Significance: Light-field microscopy (LFM) enables fast, light-efficient, volumetric imaging of neuronal activity with calcium indicators. Calcium transients differ in temporal signal-to-noise ratio (tSNR) and spatial confinement when extracted from volumes reconstructed by different algorithms. Aim: We evaluated the capabilities and limitations of two light-field reconstruction algorithms for calcium fluorescence imaging. Approach: We acquired light-field image series from neurons either bulk-labeled or filled intracellularly with the red-emitting calcium dye CaSiR-1 in acute mouse brain slices. We compared the tSNR and spatial confinement of calcium signals extracted from volumes reconstructed with synthetic refocusing and Richardson-Lucy three-dimensional deconvolution with and without total variation regularization. Results: Both synthetic refocusing and Richardson-Lucy deconvolution resolved calcium signals from single cells and neuronal dendrites in three dimensions. Increasing deconvolution iteration number improved spatial confinement but reduced tSNR compared with synthetic refocusing. Volumetric light-field imaging did not decrease calcium signal tSNR compared with interleaved, widefield image series acquired in matched planes. Conclusions: LFM enables high-volume rate, volumetric imaging of calcium transients in single cell somata (bulk-labeled) and dendrites (intracellularly loaded). The trade-offs identified for tSNR, spatial confinement, and computational cost indicate which of synthetic refocusing or deconvolution can better realize the scientific requirements of future LFM calcium imaging applications.

Light-Field Microscopy for Optical Imaging of Neuronal Activity: When Model-Based Methods Meet Data-Driven Approaches.

Understanding how networks of neurons process information is one of the key challenges in modern neuroscience. A necessary step to achieve this goal is to be able to observe the dynamics of large populations of neurons over a large area of the brain. Light-field microscopy (LFM), a type of scanless microscope, is a particularly attractive candidate for high-speed three-dimensional (3D) imaging. It captures volumetric information in a single snapshot, allowing volumetric imaging at video frame-rates. Specific features of imaging neuronal activity using LFM call for the development of novel machine learning approaches that fully exploit priors embedded in physics and optics models. Signal processing theory and wave-optics theory could play a key role in filling this gap, and contribute to novel computational methods with enhanced interpretability and generalization by integrating model-driven and data-driven approaches. This paper is devoted to a comprehensive survey to state-of-the-art of computational methods for LFM, with a focus on model-based and data-driven approaches.

Subcellular resolution three-dimensional light-field imaging with genetically encoded voltage indicators.

Significance: Light-field microscopy (LFM) enables high signal-to-noise ratio (SNR) and light efficient volume imaging at fast frame rates. Voltage imaging with genetically encoded voltage indicators (GEVIs) stands to particularly benefit from LFM's volumetric imaging capability due to high required sampling rates and limited probe brightness and functional sensitivity. Aim: We demonstrate subcellular resolution GEVI light-field imaging in acute mouse brain slices resolving dendritic voltage signals in three spatial dimensions. Approach: We imaged action potential-induced fluorescence transients in mouse brain slices sparsely expressing the GEVI VSFP-Butterfly 1.2 in wide-field microscopy (WFM) and LFM modes. We compared functional signal SNR and localization between different LFM reconstruction approaches and between LFM and WFM. Results: LFM enabled three-dimensional (3-D) localization of action potential-induced fluorescence transients in neuronal somata and dendrites. Nonregularized deconvolution decreased SNR with increased iteration number compared to synthetic refocusing but increased axial and lateral signal localization. SNR was unaffected for LFM compared to WFM. Conclusions: LFM enables 3-D localization of fluorescence transients, therefore eliminating the need for structures to lie in a single focal plane. These results demonstrate LFM's potential for studying dendritic integration and action potential propagation in three spatial dimensions.

3D Localization for Light-Field Microscopy via Convolutional Sparse Coding on Epipolar Images.

Light-field microscopy (LFM) is a type of all-optical imaging system that is able to capture 4D geometric information of light rays and can reconstruct a 3D model from a single snapshot. In this paper, we propose a new 3D localization approach to effectively detect 3D positions of neuronal cells from a single light-field image with high accuracy and outstanding robustness to light scattering. This is achieved by constructing a depth-aware dictionary and by combining it with convolutional sparse coding. Specifically, our approach includes 3 key parts: light-field calibration, depth-aware dictionary construction, and localization based on convolutional sparse coding (CSC). In the first part, an observed raw light-field image is calibrated and then decoded into a two-plane parameterized 4D format which leads to the epi-polar plane image (EPI). The second part involves simulating a set of light-fields using a wave-optics forward model for a ball-shaped volume that is located at different depths. Then, a depth-aware dictionary is constructed where each element is a synthetic EPI associated to a specific depth. Finally, by taking full advantage of the sparsity prior and shift-invariance property of EPI, 3D localization is achieved via convolutional sparse coding on an observed EPI with respect to the depth-aware EPI dictionary. We evaluate our approach on both non-scattering specimen (fluorescent beads suspended in agarose gel) and scattering media (brain tissues of genetically encoded mice). Extensive experiments demonstrate that our approach can reliably detect the 3D positions of granular targets with small Root Mean Square Error (RMSE), high robustness to optical aberration and light scattering in mammalian brain tissues.

All-optical crosstalk-free manipulation and readout of Chronos-expressing neurons.

All optical neurophysiology allows manipulation and readout of neural network activity with single-cell spatial resolution and millisecond temporal resolution. Neurons can be made to express proteins that actuate transmembrane currents upon light absorption, enabling optical control of membrane potential and action potential signalling. In addition, neurons can be genetically or synthetically labelled with fluorescent reporters of changes in intracellular calcium concentration or membrane potential. Thus, to optically manipulate and readout neural activity in parallel, two spectra are involved: the action spectrum of the actuator, and the absorption spectrum of the fluorescent reporter. Due to overlap in these spectra, previous all-optical neurophysiology paradigms have been hindered by spurious activation of neuronal activity caused by the readout light. Here, we pair the blue-green absorbing optogenetic actuator, Chronos, with a deep red-emitting fluorescent calcium reporter CaSiR-1. We show that cultured Chinese hamster ovary cells transfected with Chronos do not exhibit transmembrane currents when illuminated with wavelengths and intensities suitable for exciting one-photon CaSiR-1 fluorescence. We then demonstrate crosstalk-free, high signal-to-noise ratio CaSiR-1 red fluorescence imaging at 100 frames s-1 of Chronos-mediated calcium transients evoked in neurons with blue light pulses at rates up to 20 Hz. These results indicate that the spectral separation between red light excited fluorophores, excited efficiently at or above 640 nm, with blue-green absorbing opsins such as Chronos, is sufficient to avoid spurious opsin actuation by the imaging wavelengths and therefore enable crosstalk-free all-optical neuronal manipulation and readout.

Corrigendum: Single-Neuron Level One-Photon Voltage Imaging With Sparsely Targeted Genetically Encoded Voltage Indicators.

[This corrects the article DOI: 10.3389/fncel.2019.00039.].

Single-Neuron Level One-Photon Voltage Imaging With Sparsely Targeted Genetically Encoded Voltage Indicators.

Voltage imaging of many neurons simultaneously at single-cell resolution is hampered by the difficulty of detecting small voltage signals from overlapping neuronal processes in neural tissue. Recent advances in genetically encoded voltage indicator (GEVI) imaging have shown single-cell resolution optical voltage recordings in intact tissue through imaging naturally sparse cell classes, sparse viral expression, soma restricted expression, advanced optical systems, or a combination of these. Widespread sparse and strong transgenic GEVI expression would enable straightforward optical access to a densely occurring cell type, such as cortical pyramidal cells. Here we demonstrate that a recently described sparse transgenic expression strategy can enable single-cell resolution voltage imaging of cortical pyramidal cells in intact brain tissue without restricting expression to the soma. We also quantify the functional crosstalk in brain tissue and discuss optimal imaging rates to inform future GEVI experimental design.

High speed functional imaging with source localized multifocal two-photon microscopy.

Multifocal two-photon microscopy (MTPM) increases imaging speed over single-focus scanning by parallelizing fluorescence excitation. The imaged fluorescence's susceptibility to crosstalk, however, severely degrades contrast in scattering tissue. Here we present a source-localized MTPM scheme optimized for high speed functional fluorescence imaging in scattering mammalian brain tissue. A rastered line array of beamlets excites fluorescence imaged with a complementary metal-oxide-semiconductor (CMOS) camera. We mitigate scattering-induced crosstalk by temporally oversampling the rastered image, generating grouped images with structured illumination, and applying Richardson-Lucy deconvolution to reassign scattered photons. Single images are then retrieved with a maximum intensity projection through the deconvolved image groups. This method increased image contrast at depths up to 112 µm in scattering brain tissue and reduced functional crosstalk between pixels during neuronal calcium imaging. Source-localization did not affect signal-to-noise ratio (SNR) in densely labeled tissue under our experimental conditions. SNR decreased at low frame rates in sparsely labeled tissue, with no effect at frame rates above 50 Hz. Our non-descanned source-localized MTPM system enables high SNR, 100 Hz capture of fluorescence transients in scattering brain, increasing the scope of MTPM to faster and smaller functional signals.

Submillisecond Optogenetic Control of Neuronal Firing with Two-Photon Holographic Photoactivation of Chronos.

Optogenetic neuronal network manipulation promises to unravel a long-standing mystery in neuroscience: how does microcircuit activity relate causally to behavioral and pathological states? The challenge to evoke spikes with high spatial and temporal complexity necessitates further joint development of light-delivery approaches and custom opsins. Two-photon (2P) light-targeting strategies demonstrated in-depth generation of action potentials in photosensitive neurons both in vitro and in vivo, but thus far lack the temporal precision necessary to induce precisely timed spiking events. Here, we show that efficient current integration enabled by 2P holographic amplified laser illumination of Chronos, a highly light-sensitive and fast opsin, can evoke spikes with submillisecond precision and repeated firing up to 100 Hz in brain slices from Swiss male mice. These results pave the way for optogenetic manipulation with the spatial and temporal sophistication necessary to mimic natural microcircuit activity.SIGNIFICANCE STATEMENT To reveal causal links between neuronal activity and behavior, it is necessary to develop experimental strategies to induce spatially and temporally sophisticated perturbation of network microcircuits. Two-photon computer generated holography (2P-CGH) recently demonstrated 3D optogenetic control of selected pools of neurons with single-cell accuracy in depth in the brain. Here, we show that exciting the fast opsin Chronos with amplified laser 2P-CGH enables cellular-resolution targeting with unprecedented temporal control, driving spiking up to 100 Hz with submillisecond onset precision using low laser power densities. This system achieves a unique combination of spatial flexibility and temporal precision needed to pattern optogenetically inputs that mimic natural neuronal network activity patterns.

Advances in two photon scanning and scanless microscopy technologies for functional neural circuit imaging.

Recent years have seen substantial developments in technology for imaging neural circuits, raising the prospect of large scale imaging studies of neural populations involved in information processing, with the potential to lead to step changes in our understanding of brain function and dysfunction. In this article we will review some key recent advances: improved fluorophores for single cell resolution functional neuroimaging using a two photon microscope; improved approaches to the problem of scanning active circuits; and the prospect of scanless microscopes which overcome some of the bandwidth limitations of current imaging techniques. These advances in technology for experimental neuroscience have in themselves led to technical challenges, such as the need for the development of novel signal processing and data analysis tools in order to make the most of the new experimental tools. We review recent work in some active topics, such as region of interest segmentation algorithms capable of demixing overlapping signals, and new highly accurate algorithms for calcium transient detection. These advances motivate the development of new data analysis tools capable of dealing with spatial or spatiotemporal patterns of neural activity, that scale well with pattern size.

Vortex-free phase profiles for uniform patterning with computer-generated holography.

Computer-generated holography enables efficient light pattern generation through phase-only wavefront modulation. While perfect patterning usually requires control over both phase and amplitude, iterative Fourier transform algorithms (IFTA) can achieve phase-only approximations which maximize light efficiency at the cost of uniformity. The phase being unconstrained in the output plane, it can vary abruptly in some regions leading to destructive interferences. Among such structures phase vortices are the most common. Here we demonstrate theoretically, numerically and experimentally, a novel approach for eliminating phase vortices by spatially filtering the phase input to the IFTA, combining it with phase-based complex amplitude control at the spatial light modulator (SLM) plane to generate smooth shapes. The experimental implementation is achieved performing complex amplitude modulation with a phase-only SLM. This proposed experimental scheme offers a continuous and centered field of excitation. Lastly, we characterize achievable trade-offs between pattern uniformity, diffraction efficiency, and axial confinement.

Dendrites Enable a Robust Mechanism for Neuronal Stimulus Selectivity.

Hearing, vision, touch: underlying all of these senses is stimulus selectivity, a robust information processing operation in which cortical neurons respond more to some stimuli than to others. Previous models assume that these neurons receive the highest weighted input from an ensemble encoding the preferred stimulus, but dendrites enable other possibilities. Nonlinear dendritic processing can produce stimulus selectivity based on the spatial distribution of synapses, even if the total preferred stimulus weight does not exceed that of nonpreferred stimuli. Using a multi-subunit nonlinear model, we demonstrate that stimulus selectivity can arise from the spatial distribution of synapses. We propose this as a general mechanism for information processing by neurons possessing dendritic trees. Moreover, we show that this implementation of stimulus selectivity increases the neuron's robustness to synaptic and dendritic failure. Importantly, our model can maintain stimulus selectivity for a larger range of loss of synapses or dendrites than an equivalent linear model. We then use a layer 2/3 biophysical neuron model to show that our implementation is consistent with two recent experimental observations: (1) one can observe a mixture of selectivities in dendrites that can differ from the somatic selectivity, and (2) hyperpolarization can broaden somatic tuning without affecting dendritic tuning. Our model predicts that an initially nonselective neuron can become selective when depolarized. In addition to motivating new experiments, the model's increased robustness to synapses and dendrites loss provides a starting point for fault-resistant neuromorphic chip development.

Cortical Interneuron Subtypes Vary in Their Axonal Action Potential Properties.

The role of interneurons in cortical microcircuits is strongly influenced by their passive and active electrical properties. Although different types of interneurons exhibit unique electrophysiological properties recorded at the soma, it is not yet clear whether these differences are also manifested in other neuronal compartments. To address this question, we have used voltage-sensitive dye to image the propagation of action potentials into the fine collaterals of axons and dendrites in two of the largest cortical interneuron subtypes in the mouse: fast-spiking interneurons, which are typically basket or chandelier neurons; and somatostatin containing interneurons, which are typically regular spiking Martinotti cells. We found that fast-spiking and somatostatin-expressing interneurons differed in their electrophysiological characteristics along their entire dendrosomatoaxonal extent. The action potentials generated in the somata and axons, including axon collaterals, of somatostatin-expressing interneurons are significantly broader than those generated in the same compartments of fast-spiking inhibitory interneurons. In addition, action potentials back-propagated into the dendrites of somatostatin-expressing interneurons much more readily than fast-spiking interneurons. Pharmacological investigations suggested that axonal action potential repolarization in both cell types depends critically upon Kv1 channels, whereas the axonal and somatic action potentials of somatostatin-expressing interneurons also depend on BK Ca(2+)-activated K(+) channels. These results indicate that the two broad classes of interneurons studied here have expressly different subcellular physiological properties, allowing them to perform unique computational roles in cortical circuit operations. SIGNIFICANCE STATEMENT: Neurons in the cerebral cortex are of two major types: excitatory and inhibitory. The proper balance of excitation and inhibition in the brain is critical for its operation. Neurons contain three main compartments: dendritic, somatic, and axonal. How the neurons receive information, process it, and pass on new information depends upon how these three compartments operate. While it has long been assumed that axons are simply for conducting information from the cell body to the synapses, here we demonstrate that the axons of different types of interneurons, the inhibitory cells, possess differing electrophysiological properties. This result implies that differing types of interneurons perform different tasks in the cortex, not only through their anatomical connections, but also through how their axons operate.

Computer-generated holography enhances voltage dye fluorescence discrimination in adjacent neuronal structures.

Voltage-sensitive fluorescence indicators enable tracking neuronal electrical signals simultaneously in multiple neurons or neuronal subcompartments difficult to access with patch electrodes. However, efficient widefield epifluorescence detection of rapid voltage fluorescence transients necessitates that imaged cells and structures lie sufficiently far from other labeled structures to avoid contamination from out of focal plane and scattered light. We overcame this limitation by exciting dye fluorescence with one-photon computer-generated holography shapes contoured to axons or dendrites of interest, enabling widefield detection of voltage fluorescence with high spatial specificity. By shaping light onto neighboring axons and dendrites, we observed that dendritic back-propagating action potentials were broader and slowly rising compared with axonal action potentials, differences not measured in the same structures illuminated with a large "pseudowidefield" (pWF) spot of the same excitation density. Shaped illumination trials showed reduced baseline fluorescence, higher baseline noise, and fractional fluorescence transient amplitudes two times greater than trials acquired with pWF illumination of the same regions.

Somatic membrane potential and Kv1 channels control spike repolarization in cortical axon collaterals and presynaptic boutons.

The shape of action potentials invading presynaptic terminals, which can vary significantly from spike waveforms recorded at the soma, may critically influence the probability of synaptic neurotransmitter release. Revealing the conductances that determine spike shape in presynaptic boutons is important for understanding how changes in the electrochemical context in which a spike is generated, such as subthreshold depolarization spreading from the soma, can modulate synaptic strength. Utilizing recent improvements in the signal-to-noise ratio of voltage-sensitive dye imaging in mouse brain slices, we demonstrate that intracortical axon collaterals and en passant presynaptic terminals of layer 5 pyramidal cells exhibit a high density of Kv1 subunit-containing ion channels, which generate a slowly inactivating K(+) current critically important for spike repolarization in these compartments. Blockade of the current by low doses of 4-aminopyridine or α-dendrotoxin dramatically slows the falling phase of action potentials in axon collaterals and presynaptic boutons. Furthermore, subthreshold depolarization of the soma broadened action potentials in collaterals bearing presynaptic boutons, an effect abolished by blocking Kv1 channels with α-dendrotoxin. These results indicate that action potential-induced synaptic transmission may operate through a mix of analog-digital transmission owing to the properties of Kv1 channels in axon collaterals and presynaptic boutons.

The spatio-temporal characteristics of action potential initiation in layer 5 pyramidal neurons: a voltage imaging study.

The spatial pattern of Na(+) channel clustering in the axon initial segment (AIS) plays a critical role in tuning neuronal computations, and changes in Na(+) channel distribution have been shown to mediate novel forms of neuronal plasticity in the axon. However, immunocytochemical data on channel distribution may not directly predict spatio-temporal characteristics of action potential initiation, and prior electrophysiological measures are either indirect (extracellular) or lack sufficient spatial resolution (intracellular) to directly characterize the spike trigger zone (TZ). We took advantage of a critical methodological improvement in the high sensitivity membrane potential imaging (V(m) imaging) technique to directly determine the location and length of the spike TZ as defined in functional terms. The results show that in mature axons of mouse cortical layer 5 pyramidal cells, action potentials initiate in a region ∼20 µm in length centred between 20 and 40 µm from the soma. From this region, the AP depolarizing wave invades initial nodes of Ranvier within a fraction of a millisecond and propagates in a saltatory fashion into axonal collaterals without failure at all physiologically relevant frequencies. We further demonstrate that, in contrast to the saltatory conduction in mature axons, AP propagation is non-saltatory (monotonic) in immature axons prior to myelination.

State-dependent auditory evoked hemodynamic responses recorded optically with indwelling photodiodes.

Implantable optical technologies provide measurements of cerebral hemodynamic activity from freely behaving animals without movement constraint or anesthesia. In order to study state-dependent neural evoked responses and the consequential hemodynamic response, we simultaneously measured EEG and scattered light changes in chronically implanted rats. Recordings took place under freely behaving conditions, allowing us to compare the evoked responses across wake, sleep, and anesthetized states. The largest evoked electrical and optical responses occurred during quiet sleep compared to wake and REM sleep, while isoflurane anesthesia showed a large, late burst of electrical activity synchronized to the stimulus but an earlier optical response.

In vitro and in vivo noise analysis for optical neural recording.

Laser diodes (LD) are commonly used for optical neural recordings in chronically recorded animals and humans, primarily due to their brightness and small size. However, noise introduced by LDs may counteract the benefits of brightness when compared to low-noise light-emitting diodes (LEDs). To understand noise sources in optical recordings, we systematically compared instrument and physiological noise profiles in two recording paradigms. A better understanding of noise sources can help improve optical recordings and make them more practical with fewer averages. We stimulated lobster nerves and a rat cortex, then compared the root mean square (RMS) noise and signal-to-noise ratios (SNRs) of data obtained with LED, superluminescent diode (SLD), and LD illumination for different numbers of averages. The LED data exhibited significantly higher SNRs in fewer averages than LD data in all recordings. In the absence of tissue, LED noise increased linearly with intensity, while LD noise increased sharply in the transition to lasing and settled to noise levels significantly higher than the LED's, suggesting that speckle noise contributed to the LD's higher noise and lower SNRs. Our data recommend low coherence and portable light sources for in vivo chronic neural recording applications.