Reviewer #2 (Public review):
Summary:
This manuscript presents a miniaturized microscope capable of structured illumination using the HiLo approach. This is achieved through adding a second excitation path with a Ronchi grating to generate time-multiplexed excitation that sequentially generates wide-field and lined images at 15 FPS each. Through lightweight HiLo computation, this results in a 15 FPS recording with larger spatial components removed. Structured illumination can reduce background signal, out-of-focus neuronal signals, and can achieve optical sectioning for multi-plane imaging in some cases. The core engineering idea in this manuscript is clever and motivated by a fundamentally limiting aspect of wide-field imaging.
My central problem is that I cannot identify what the authors are arguing for overall. The paper evaluates HiLo against two conditions: raw widefield ROI pixel-averaging, and full CNMF source extraction. Nothing in between the 2 extremes is considered but this is the space where it seems the real gains from this approach could be made. Within those two conditions, the authors report that after CNMF extraction, HiLo provides no significant advantage in SNR or decoding accuracy. A spatial-information difference remains, but it disappears for deconvolved signals after event-rate matching. The conclusion I can draw from the manuscript's central figure is that one should use a conventional widefield miniscope and run CNMF. The remaining claim, that HiLo lets you skip offline demixing, is a narrow one, and it is not demonstrated, including in the authors' closed-loop experiment, where spatial footprints were identified offline in advance. I do not think this is what the authors intend to convey, but the manuscript lacks support for another story.
Beyond this, the manuscript has a number of unsupported critical claims and gaps in methodology and comparisons, which I try to outline below. I think the approach and tool itself are interesting, but as presented, I am not able to judge how much it improves real-world 1p imaging, and I suspect the improvement is smaller and more specific than the framing implies. To be clear, I think there is value in this work, but I think the claims need to be narrowed to what is currently demonstrated or analysis be added to support unsupported claims.
Major Concerns:
(1) The comparison space is missing its middle, where HiLo would be most relevant to compare against other widefield processing
HiLo combines a physical change in hardware and acquisition with a computational processing step, and the manuscript never separates the two. The demonstrated advantages of this approach are either measured against unprocessed widefield or mostly disappear once CNMF is applied.
The comparison that likely is most relevant, widefield with lightweight preprocessing alone, spatial filtering and background subtraction as simple as removing a minimum projection, appears nowhere. These steps are standard in the field and are built into essentially every CNMF pipeline, including Minian, MIN1PIPE, and CaImAn. As presented, it is unclear whether the physical Ronchi mask achieves anything that lightweight preprocessing cannot.
The current comparisons do not establish how much benefit specifically requires patterned acquisition rather than lightweight digital background suppression: HiLo + CNMF performs no better than widefield + CNMF on SNR or decoding accuracy. The small spatial information difference that remains does not survive the authors' own activity rate matching control. If optical sectioning were removing background that computation could not, these comparisons should favor HiLo.
Separately, and I do not see this acknowledged anywhere, HiLo is given twice the raw data, since each output frame is built from two camera frames while the widefield condition is one of those streams at 15 FPS.
(2) Inconsistent use of CNMF-E comparisons
CNMF is run with different parameters on the two modalities, justified on the grounds that the modalities "differ in signal intensity and signal-to-noise characteristics." That difference is the quantity being measured, so the justification is possibly circular.
CNMF is also used in two structurally different roles that are never distinguished. In one, it serves only to find and segment cell footprints. In the other, its extracted traces are used directly. The first use discards demixing, which is a major point of CNMF and the reason it is standard in the field, not SNR improvement. The manuscript does not address demixing, which is arguably the most important processing challenge in 1P imaging due to scattering and cell overlap. Optical sectioning can indeed help in certain cases of this but that isn't argued, supported, or shown here. It leaves me asking what I think is an important question given how this manuscript is structured: what are the limits on how HiLo can separate out-of-plane cells?
(3) Concerns around simulated data
The performance of HiLo depends on the patterned line spacing/frequency relative to the spatial scale of features being suppressed vs retained. The simulated data does not address this dependency. The grating period is given as roughly 28 µm at the sample plane, but the spatial frequency response it produces is never characterized, and the simulated neurons and local background are described only as having "distinct dimensions" with no sizes stated. The one ratio that determines the outcome is therefore unspecified on both sides. Without a spatial frequency filtering curve, or performance across a range of spatial scales, the simulation could simply be tuned to the pattern spacing that gives optimal results. Related to this: how was the pattern line spacing determined? Was it optimized for the brain region, microscope parameters, expression pattern, and imaging depth used in experiments? This is not discussed.
(4) Support for calling HiLo "multi-plane or volumetric imaging"
I do not think the data support this being two distinct planes in the brain. The quoted axial sectioning is narrower than the stated separation between the two planes, which should leave little signal leaking between them, yet Figure 3D reports a substantial fraction of cells detected in both. These are difficult to reconcile.
Field curvature is a possible straightforward alternative explanation for two focal settings yielding largely non-overlapping cell sets, and nothing in Figure 3 distinguishes it from true depth sectioning. It would equally explain the SNR difference in Figure 3E as increased aberration toward the edge of the GRIN lens. CA1 is also approximately a single cell layer thick, which makes resolving cells at truly different depths here unlikely. I am not suggesting that these alternatives necessarily explain the result, but they should be distinguished from true depth-resolved optical sectioning before the data are characterized as volumetric imaging.
Lastly, the claim in lines 165-168 that this "increases neuronal yield while maintaining reliable signal quality across imaging depths" lacks the proper comparison. Figure 3 has no widefield condition, so yield is compared only against a single HiLo plane. The proper comparison would be against processed widefield imaging at a single optimally-chosen focal plane. Lower per-plane counts are also expected by design under optical sectioning and cannot serve as evidence of improved performance without that control.
(5) Place cell analysis
I have several methodological concerns with the place cell analysis:
a) Place field analysis should use deconvolved neural activity, not calcium signals. Deconvolved signals appear only in Figure 4 and Figure Supplement 2; the main figure uses ΔF/F for the ROI condition and raw CNMF fluorescence for the other.
b) Lines 170-172 motivate this analysis by the "improved performance of HiLo imaging in suppressing background fluorescence for ROI-based signals," but ROI averaging without demixing is not the standard analysis for place-cell inference. CNMF or similar demixing is generally used and should be the main comparison.
Concerning the spatial info content differences in Figure 4: how do the authors know that HiLo+CNMF is more correct than WF+CNMF? They observe a minor but significant difference, but there is no literature (and none cited) that would support the implied conclusion that, since HiLo results in higher spatial info content, it is somehow extracting more accurate neuronal dynamics. Even if decoding performance improved, Figure 4I shows no significance for CNMF-extracted signals, and artificially suppressing weak parts of a place field could improve these metrics without reflecting truer neural signals. The authors' own data argue for exactly this: HiLo signals show significantly lower event rates, and when event rates are matched between conditions the spatial information difference disappears (Figure 4-figure supplement 2). However, this control was applied only to the deconvolved signals and not to the ROI-based signals that carry the central claim.
(6) Real-time experiment adds little value
The closed-loop demonstration does not establish that HiLo enables anything beyond current capabilities. Spatial footprints are identified offline before training, and there are existing real-time streaming calcium imaging approaches that work once cell footprints are identified. The DeCalciOn approach (Chen et al., eLife 2023) demonstrates that similar decoder performance can be achieved without even requiring cell ROIs.
I do note that the discussion states these demonstrations are not intended to define the primary application domain of the device. Nevertheless, the title leads with real-time neural decoding, and the demonstration occupies a main figure. I don't believe anything in it demonstrates that the described experiments are enabled by HiLo and could not be done with existing widefield systems right now. The authors would benefit from being able to directly show and argue for this, or they should change the framing of this section.
Overall Assessment:
The core engineering contribution, integrating structured illumination into a miniature microscope with a reasonably direct approach, is a clever design that is worth reporting. However, the manuscript's central claims about improved performance are undermined by how the comparisons are set up. Fully processed HiLo data is placed against either raw widefield or full CNMF, with nothing in between, and that middle ground is where the standard lightweight processing used across the field actually sits. Where CNMF is applied to both, the advantage largely disappears, which makes it difficult to tell what the physical approach is contributing at all. The multi-plane imaging claims may be overstated and potentially confounded by field curvature, while the correspondence between the bench calibration and effective in vivo axial sectioning is not directly validated. The place cell analysis has methodological issues, and the real-time decoding demonstration does not establish that HiLo enables capabilities beyond what existing systems can achieve.
My view is that the claims need to be narrowed to what is currently demonstrated, or the analyses added to support them, and I would be happy to read and review a revised version.