A structured-illumination miniscope for optically sectioned imaging and real-time neural decoding
nvidia/Gemma-4-26B-A4B-NVFP4 · academic_accessible/eval 94%/5 min read/Jul 25, 2026
Generated by a local model (nvidia/Gemma-4-26B-A4B-NVFP4) from a scientific paper, claim-checked against the full text. Provenance is open by design.
Can We See Through the Brain's Fog?
Researchers want to watch the brain in action while animals move naturally. However, the view is often obscured by a biological haze. Standard miniscopes—tiny, wearable microscopes—allow scientists to track calcium signaling (a proxy for neuronal activity) in freely behaving mice. But these devices suffer from a fundamental clarity problem. They capture light from every layer of the tissue at once. This creates a blurry background of out-of-focus fluorescence (light emitted from molecules not in the focal plane) that masks the neurons scientists want to study.
A new study introduces a lightweight (<3 g), low-cost solution to this "fog" using structured illumination. Instead of flooding the brain with uniform light, the researchers implement a technique called HiLo microscopy. This achieves optical sectioning (the ability to selectively image a thin slice of tissue). It rejects the glow from everything above and below the target plane. This approach aims to bridge the gap between cheap, blurry widefield microscopes and expensive, complex multiphoton systems.
Solving the out-of-focus dilemma
The central question driving this work is whether a miniaturized, single-photon system can achieve high-contrast optical sectioning. This is a feature typically reserved for much larger, more expensive laboratory microscopes. Specifically, the authors sought to suppress background fluorescence at the moment of acquisition. They wanted to do this without sacrificing the portability required for studying animals in motion.
In traditional widefield miniscopes, the excitation light hits the entire volume of the tissue. This causes neurons at various depths to glow simultaneously. This creates a massive amount of "out-of-focus" background signal. This background acts like a heavy fog. It degrades the signal-to-noise ratio (SNR, a measure of signal strength relative to background noise) and makes it difficult to distinguish single-neuron flashes from the collective glow.
Moving beyond post-hoc cleaning
Until now, the field has largely relied on two strategies to deal with this blur. The first is computational "demixing" (mathematically separating overlapping signals). This uses complex algorithms like CNMF-E to untangle signals after the data has been recorded. The second is multiphoton microscopy. This uses specialized lasers to excite only a tiny point at a specific depth, providing inherent sectioning.
The authors highlight significant trade-offs in these approaches. Computational cleaning is a "post hoc" process. It happens after the fact and cannot recover information lost to noise during the initial recording. These algorithms are also computationally intensive. This creates a bottleneck for real-time applications. Conversely, multiphoton systems are heavy and power-hungry. They are also technically difficult to stabilize on a moving animal. The researchers aimed for a middle ground. They sought to improve image quality at the source using a simple, lightweight hardware modification.
Implementing the HiLo architecture
The researchers developed a HiLo miniscope using a Ronchi grating (a simple optical element that creates a striped pattern of light). As shown in [Figure 1A] and [Figure 1B], the device weighs less than 3 grams. The system operates through time-multiplexed excitation. It alternates between a uniform light pattern and a structured, striped pattern.
By comparing these two types of images, the HiLo reconstruction algorithm separates the in-focus signal from the diffuse background. The authors validated this using a standardized resolution target. They reported a lateral resolution of 256 line pairs per mm [Figure 1E]. They also achieved an axial full width at half maximum (FWHM, the width of a signal at half its peak intensity) of 23 $\mu$m [Figure 1—figure supplement 1C]. This axial measurement represents the thickness of the "slice" the microscope can clearly see. To expand the imaging range, they integrated an electrowetting liquid lens. This allows users to change the focal plane by adjusting voltage [Figure 3A].
Higher fidelity from simpler signals
The study finds that this acquisition-level suppression fundamentally changes the utility of the data. In simulations, the HiLo method recovered neural signals with much higher fidelity than conventional widefield imaging [Figure 2C, Figure 2D].
When tested in vivo on the hippocampal CA1 region of mice, the results were notable. The authors report that simple ROI-averaged signals—which average light within a predefined area around a neuron—approached the performance of complex, algorithmically-extracted signals [Figure 2G]. This suggests that because the "fog" is removed during imaging, researchers may not always need heavy-duty math for clean data.
Furthermore, the HiLo miniscope demonstrated superior spatial information readout. In tests involving animal position decoding, the errors were significantly lower when using HiLo-derived signals compared to widefield data [Figure 4I]. The researchers also proved the system could support a closed-loop brain-machine interface (BMI). They showed that real-time decoding of population activity could successfully trigger rewards in mice .
Figure 5. Closed-loop hippocampal BMI training. (A) Experimental setup for a closed-loop BMI paradigm. The real-time decoding algorithm compares ongoing population vector activity to a pre-established odor-evoked template. If the correlation exceeds a threshold, a reward is triggered. Training consists of acquisition, extinction, and reacquisition phases, with rewards delivered during acquisition and reacquisition but not extinction. (B) Example session of real-time correlation of population vector activity with the template during training. Red dashed line represents correlation threshold. Vertical blue lines are reward delivery and gray bars are licks. (C) Threshold-crossing events (% of total) across the acquisition process. Gray lines indicate individual animals, with the bolded gray line showing the example mouse in (B), (E)-(H); green line shows the population average. Correlation between acquisition progress and threshold-crossing events is tested using Spearman correlation. (D) Frequency of crossing events per minute during the acquisition, extinction, and reacquisition phases. Bolded gray line corresponds to the example mouse. Paired lines indicate within-session comparisons; boxes show median and interquartile range; whiskers indicate extreme values. (E) Example behavioral session showing lick rasters aligned to threshold-crossing events (dashed line); shaded regions indicate training blocks. (F) Lick rate as a function of time relative to threshold-crossing events for the example session. (G) Lick rate preceding threshold-crossing events ( -2 to -1 s), corrected by subtracting baseline lick rate ( -4 to -3 s), across acquisition, extinction, and reacquisition blocks for the same session. Statistical significance was assessed using rank sum test (* 𝑝 < 0 . 05 , ** 𝑝 < 0 . 01 ; n.s., not significant). (H) Baseline ( -6 to -3 s) and peri-crossing ( -1 to 2 s) lick rates during extinction. Statistical significance was assessed using a paired 𝑡 -test ( 𝑛 = 32 crossing events; **** 𝑝 < 0 . 0001 ). In (G) and (H), colored dots represent individual crossing events; white dots indicate medians.
Bridging the gap in neurotechnology
This work suggests a shift in how we approach large-scale neural recordings. If this technology generalizes, it provides a "third way" for neuroscientists. It offers the accessibility of widefield imaging with a clarity approaching multiphoton systems.
For researchers focused on real-time interaction, such as closed-loop studies, this device removes computational latency. It avoids the delay caused by heavy post-processing. Additionally, multi-plane imaging with a liquid lens [Figure 3B] allows researchers to increase their "neuronal yield" (the total number of cells recorded). This increases sampling without needing entirely new systems.
However, the authors note certain trade-offs. The HiLo miniscope does not offer the same depth of tissue penetration or scattering rejection as multiphoton microscopy. It also does not explicitly separate two physically overlapping neurons. It is a tool for clarity, not a complete replacement for advanced computational demixing. It serves as a complementary solution for high-quality population recordings.
Figures from the paper
Figure 1. Design and performance of the HiLo miniscope. (A) Cross-sectional schematic of the HiLo miniscope. (B) Dimensions of the HiLo miniscope. (C) Time-division multiplexing scheme used to acquire uniformly illuminated and structured illumination images. The total acquisition frame rate is 30 frames per second. (D) Uniformly illuminated image and corresponding structured illumination image of a thin, uniform fluorescent plane. The Ronchi grating has a spatial period of approximately 28 𝜇 mat the sample plane. (E) Uniformly illuminated image of a USAF 1951 resolution target. Inset shows a magnified view of the highlighted region; the finest resolvable feature corresponds to 256 lp/mm (Group 8, Element 1). (F) Comparison of uniformly illuminated widefield imaging and optical-sectioned HiLo miniscope imaging in a Thy1-GFP mouse brain slice. Intensity profiles along the blue and magenta line cuts illustrate the reduction of background fluorescence and the enhancement of image contrast achieved by the HiLo miniscope.Figure 2. HiLo imaging improves signal fidelity and ROI-based calcium signal quality. (A) Schematic of simulated data generation. Each synthetic widefield frame was constructed as a weighted sum of three components: an in-focus neural activity signal ("true signal"), a spatially localized out-of-focus background ("local background"), and a global background signal ("global background"). (B) HiLo reconstruction of the simulated widefield image shown in (A). (C) Example fluorescence traces extracted from the simulated ROI indicated in (A) and (B). (D) Quantitative comparison of signal fidelity in simulations. Scatter plot shows the correlation coefficient between extracted signals and ground-truth activity for widefield imaging versus HiLo imaging. Each dot represents one simulated ROI; the green dot indicates the example ROI shown in (C). (E) In vivo calcium imaging from hippocampal CA1 in freely moving mice. Rows show raw widefield images, raw HiLo images, and CNMF-E-denoised signals. Left, maximum-intensity projections (MIPs) over 1000 frames; middle, representative single frames with example ROIs outlined; right, representative fluorescence traces from the same color-coded ROIs. (F) Comparison of mean Δ 𝐹 ∕ 𝐹 between widefield and HiLo ROI signals across all ROIs ( 𝑛 = 1 , 513 ROIs from 5 mice). Each point represents one ROI. (G) Signal-to-noise ratio comparison across extraction methods, including ROI pixel-averaged signals from widefield and HiLo images and CNMF-E-extracted signals from widefield and HiLo data. Colored dots represent individual ROIs; white dots indicate medians; lines connect session-averaged values from individual mice. (H) Similarity between ROI-based and CNMF-E-extracted signals. Scatter plot shows the correlation between ROI-derived fluorescence traces from widefield or HiLo images and denoised CNMF-E fluorescence traces extracted from the corresponding widefield dataset. Black dots represent individual ROIs, and colored dots indicate session means. Statistical significance in (G and H) was assessed using paired 𝑡 -tests on mouse-level mean values ( 𝑛 = 5 mice; * 𝑝 < 0 . 05 , *** 𝑝 < 0 . 001 ; n.s., not significant).Figure 3. Optical-sectioned multi-plane imaging increases the yield of imaged neurons. (A) Color-coded composite image sampled from a Thy1-GFP mouse brain slice across multiple focal depths, illustrating volumetric coverage enabled by optical-sectioned HiLo imaging. Color indicates relative imaging depth. Scale bar, 100 𝜇 m. (B) Schematic and representative maximum-intensity projection (MIP) images illustrating in vivo two-plane imaging. Example fluorescence images acquired at depths of 179 𝜇 mand 212 𝜇 mare shown, together with detected neurons color-coded by imaging plane and overlap (magenta, 212 𝜇 m; green, 179 𝜇 m; white, neurons detected in both planes). (C) Representative calcium activity traces ( Δ 𝐹 ∕ 𝐹 ) from neurons detected exclusively in the deeper plane (212 𝜇 m), exclusively in the shallower plane (179 𝜇 m), and in both planes (overlap), demonstrating reliable signal extraction across imaging depths. (D) Percentage of neurons detected in both imaging planes relative to the total number of neurons identified across planes. (E) Comparison of mean signal-to-noise ratio (SNR) between neurons detected in the deeper and shallower imaging planes. (F) Number of neurons detected in the deeper plane, the shallower plane, and after combining neurons from both planes. In (D)-(F), each dot represents one imaging session. In (E) and (F), paired lines indicate within-session comparisons. Boxes represent the median and interquartile range, and whiskers indicate the minimum and maximum values. Statistical significance was assessed using paired 𝑡 -tests ( 𝑛 = 6 mice; * 𝑝 < 0 . 05 , ** 𝑝 < 0 . 01 , *** 𝑝 < 0 . 001 ).Figure 4. HiLo imaging enhances ROI-based spatial readout from hippocampal CA1. (A) Representative matched ROI from dorsal hippocampal CA1. Left: animal trajectory (gray) overlaid with calcium event locations (red); only events exceeding 3 standard deviations of Δ 𝐹 ∕ 𝐹 are shown. Right: corresponding spatial tuning map, computed with 2.5 × 2.5 cm spatial bins and smoothed with a Gaussian kernel ( 𝜎 = 3 . 5 cm). (B) Cumulative distributions of spatial information for individually extracted ROI pixel-averaged signals from widefield and HiLo images. (C) Comparison of spatial information for matched ROIs identified in both widefield and HiLo images; each dot represents one ROI. (D) Summary of mean spatial information for matched ROIs across imaging modalities. (E)-(G) Same analyses as in (B)-(D), performed on signals extracted using the CNMF-E algorithm. (H) Bayesian decoding of animal position using ROI pixel-averaged signals from widefield and HiLo images. Actual trajectories are shown in black, and decoded positions are overlaid as colored dots for each imaging modality. (I) Decoding performance quantified as median decoding error. (J) Same decoding analyses as in (H) using CNMF-E-extracted signals. In (D), (G), and (I), paired lines indicate within-session comparisons; boxes show median and interquartile range; whiskers indicate extreme values; statistical significance was assessed using paired 𝑡 -tests ( 𝑛 = 5 mice; * 𝑝 < 0 . 05 , ** 𝑝 < 0 . 01 , **** 𝑝 < 0 . 0001 ; n.s., not significant).Figure 1-figure supplement 1. Lateral and axial resolution of the HiLo miniscope. (A) Representative image of sub-resolution fluorescent beads, 1 𝜇 m in diameter, used to characterize optical resolution. The inset shows a magnified view of a single bead. Scale bar, 100 𝜇 m. (B) Normalized lateral intensity profile along the x-axis, centered on the peak intensity of sub-resolution fluorescent beads. The Gaussian fit (blue line) yielded a lateral full width at half maximum (FWHM) of 3.1 𝜇 m. Data are presented as mean ± s.d.; n = 8 beads. (C) Axial sectioning curve measured from a thin, uniform fluorescent plane as a function of defocus. The normalized axial intensity profile (black dots) and Gaussian fit (blue line) yielded an axial FWHM of 23 𝜇 m.