Nice work! Pretty cool to see that solid execution on a relatively simple idea (auto copy/pasting single-animal blobs to synthesize overlap) can give such strong results. I'm testing AMADEUS on some C. elegans crawling data right now, and so far it handles detection and tracking of overlapping worms better than anything else I've tried.

I'm curious whether you tried or considered YOLO-seg in place of the OBB detector. Since you already have the single-animal blob masks during synthesis, it seems like they could double as training targets, and I'd guess that might help with strongly curved or coiled animals where an OBB is a pretty rough description of the body (like my worms). Is segmentation something you're thinking about for a future version? And if not, why not? Was it too difficult defining mask targets under heavy occlusion, fitting it in with the direction-class scheme, speed, or something else?

Thanks for releasing the code and the GUI!

Hi, really enjoyed this preprint. One question on the "return to arena" finding: could this reflect area-restricted search (worms tending to loop back toward any recently-vacated patch, independent of whether it was rewarding) rather than touch itself being rewarding? Since mec-4-expressing neurons could plausibly contribute to the local-search motor program rather than to a hedonic touch signal, it seems hard to fully distinguish the two accounts from the current data. Would it be feasible to run a control in which worms spend an equivalent amount of time in a flat, particle-free patch before removal, and compare their return rate to that patch? This could help clarify whether the touch-seeking interpretation is specific to particle contact or reflects a more general post-departure search behavior.

Really enjoyed this — the range of demos, from phantoms to whole Xenopus embryos, was pretty compelling to see!

One thing I was wondering about wrt the phantom characterization is that it only tests resolution at a single scattering strength. Since the biological samples afterward span a much wider scattering mean free path, it seems like it'd add a lot to turn that into an actual sweep: same phantom, more tape layers --> FWHM plotted against ℓₛ.

Also noticed the skin/VML numbers (Fig 3n,o) land pretty close to the 1 ℓₛ phantom result, which hints performance might hold up flat across a decent scattering range — but that's more a hunch from comparing separate experiments than something directly demonstrated. Curious if this is something you're planning to add, or if you already have data that speaks to it.

Hello, thanks for sharing this work -- your darkfield images are impressive for such an inexpensive setup and it's awesome how open you've been with all the files, BOM, and instructions. Makes it super easy to actually build on.

I was wondering if you could confirm the effective pixel size at the sample plane for the 12mm and 16mm lens setups? Knowing the IMX477 chip is 1.55 µm and you're capturing the full 90mm dish at 4056×3040, I'm guessing it's something like ~20-30 µm/pixel, but would love the actual number.

Thanks for putting this together!

Hi, nice work! The contrast in the reconstructed FLIM images is pretty striking, and it's nice to see it demonstrated in cells, where the actin filaments and mitochondrial structures come through so cleanly.

I had one question that I was curious about: how sensitive is the reconstruction to the choice of the four regularization parameters? And relatedly, how generalizable do you imagine these parameters to be when moving to new sample types? Just wondering if you have a sense of how much they might need to be revisited for samples with, say, very different structural complexity or signal levels.

Hello, really nice work! The channel-adaptive variant is a neat way to get cross-dataset generalization. One thing I was curious about is how many (and which) landmark channels actually matter for prediction accuracy?

Figure S11 shows that it works qualitatively with a single channel, and OpenCell is quantitative but changes both channel count and imaging domain at once, so I couldn't tell how much is lost by dropping channels alone. Seems like it should cost something real given that each channel carries independent info, and the model only sees cell identity/state through landmark morphology, so fewer channels means less to condition on. The Vermeer-XL CA vs. fixed gap in Tables S2–S4 hints at this too. A quick within-HPA ablation (nucleus only vs. + microtubule vs. + ER, same metrics) would isolate it and tell people how much fidelity they give up when they've only got, say, a Hoechst stain. Thanks btw for sharing the code and weights!

Hi, really enjoyed this preprint -- we also have a STELLARIS 8 in the lab and are looking forward to trying out your napari plugin on our own data. Thanks for developing and open-sourcing it!

Quick question about the lifetime unmixing workflow: here you mention manually selecting an ROI on the phasor plot to define the mCherry lifetime signature, then applying that fixed selection across datasets. I'm curious what your thoughts are on automating that selection step, e.g. with some kind of clustering or fitting approach directly on the phasor plot. Do you think that's feasible, or are there practical reasons it needs a human in the loop? I could see how maybe the clusters are too noisy or context-dependent to reliably pick out automatically. I ask because it'd be great to see this plugged into more hands-off analysis pipelines, but I could also see arguments for why manual curation is important here. Would love to hear your take either way!

Hello, I really enjoyed this paper, the autoencoder-based unmixing approach is elegant and the subcellular resolution results are impressive! I do have a question though about the endmember characterization: you mention here that peak assignments are based on characteristic vibrations in the fingerprint region, but the model is trained on the full spectrum including the high wavenumber region (or at least it appears that way). When you look at the HWN features in the derived endmembers -- particularly the C-H stretching bands around 2850 cm⁻¹ -- do these signatures reinforce or complicate your assignments?

In principle, the 2850/2880 ratio in the HWN region could say something about chain saturation or order that isn't really accessible from the fingerprint alone. Was it a deliberate choice to focus interpretation on the fingerprint region, or is the HWN just noisier/less informative in your data?

Nice work! This seems like a useful tool for making image analysis more accessible.

I was curious whether there was a specific reason for the limitation you mention here: "The pipeline does not currently support batch parallelization, processing images sequentially."

My first thought was that segmentation might be the main constraint (especially when using Cellpose). But it seems like one possible architecture would be to run segmentation first and save masks, then parallelize the downstream steps (puncta detection, colocalization, morphology measurements, etc.) across images or cells.

That part of the pipeline seems like it could be relatively straightforward for Claude to parallelize with something like multiprocessing or joblib and might significantly improve throughput for larger datasets, since downstream analysis would no longer be tied to per-image segmentation.

Was something like this considered, or are there constraints (e.g., Cellpose behavior, I/O, memory usage) that make parallelization less practical?

Hello! I really appreciated this work. The parameter–sensitivity analysis is especially nice for connecting the experimental results to elucidate the role of mother–daughter asymmetry.

One question I had: you mention that the cln3Δwhi5Δ mutant shows reduced mother–daughter asymmetry compared to what the pure timer model predicts, which seems like an interesting discrepancy. Do you have any hypotheses about what biological coordination or parameter correlations might explain this?

I also wondered whether the microscopy data used for the mother–daughter asymmetry analysis might be made publicly available? I think it would be a valuable resource for others in the field.

Great work on mapping how yeast organelles respond to stress — the breadth of your imaging survey is impressive! Have you considered quantifying some of these changes (e.g., % of cells with fragmented mitochondria, puncta counts, intensity ratios)? It would be especially interesting to quantify, for instance, the vacuole fusion you observed, as it might provide a more thorough account of how this varies across strains.

Additionally, making the image data (raw or processed) publicly available would be a valuable resource for others who may want to analyze or build upon your work.

Would you be willing to share the script you used to download/curate the Zenodo datasets, or provide direct references/links to the specific datasets you pulled from?

Hi — really nice work on this paper, the approach looks very promising!

I wanted to clarify one point: do the datasets you've curated already include the worm background masks, or is generating these masks part of your preprocessing pipeline? Here you mention refining the masks with connected components and morphological operations, but it’s not clear to me whether you start from raw frames or from pre-segmented images.

Thanks for clarifying!

Hi all, really enjoyed this preprint!

I think MicroSplit is a creative and elegant approach to navigating the photon budget limitation. The limitation I'm highlighting here, however, stood out to me as particularly tricky to contend with. In my experience, it's fairly common to see multi-fold brightness variation of a given structure in a single image—let alone across a dataset—due to factors like: * Differential uptake or expression of dyes or fluorescent proteins * Variability in labeling efficiency or target accessibility * Cell-to-cell heterogeneity in biological state (e.g., membrane permeability, metabolic activity)

These sources of variability seem like they could complicate MicroSplit's unmixing inferences. Do you have any thoughts on how to handle this kind of within-class brightness heterogeneity or on ways the method might be adapted to be more robust to it?

Thanks again for sharing such a neat piece of work!

Is this a gut feeling, an empirical estimate, or rooted in theory?

Thanks for sharing this work! This appears to be a super practical solution to a pervasive challenge in biology labs. I think you've identified some key issues with how a lot of research is done --- particularly the failures to document relationships between research objects and the sporadic nature of data recording. I have two questions about the practical aspects:

First, what's your experience been with researcher adoption? Are scientists you've approached generally eager to start using your platform, or do you find that you need to persuade them or provide incentives to get them on board?

Second, I'm curious how you've considered version control for protocols. In practice, it seems quite common for small tweaks to be made at various steps by different researchers, such that the "same" protocol might accumulate a number of tiny discrepancies that add up to something more significant. There's also the issue that protocols often evolve over months or years as methods are refined. How does your system handle tracking these variations and ensuring reproducibility when protocols undergo iterative changes?

I found your work on LPBT correction for pile-up distortions very impressive. I'm curious if you could propose a straightforward experiment to determine whether a commercial FLIM system is significantly affected by pile-up distortions? Perhaps using your fluorescein with potassium iodide approach as a standardized test?

I'm impressed by your open-source Python package for automated cell phenotyping from ptychography time-lapse videos, particularly your integration of Cellpose and TrackMate. That said, I'm curious about your decision to use proprietary software for preprocessing the image data.

I'm not sure what other algorithms are involved in your preprocessing workflow, but for a pure Python implementation, scikit-image provides a rolling ball background subtraction algorithm through the skimage.restoration.rolling_ball() function. Moving your preprocessing steps to Python would keep everything in the same ecosystem as the rest of your pipeline – seems like a natural fit with what you've already built.

Very nice results! I have a question about the confocal microscopy images in Figure 2c. I noticed there are a few cells visible in the DAPI channel (particularly in the top row) that show minimal or no discernible APC or GFP signal. Are these representative of the small percentage (<3%) of CAR-expressing cells that weren't successfully stained by the multimers, or are they potentially CAR-negative cells included in the sample? I'm curious about how these unstained cells relate to the high specificity rates you reported.

Are you able to tune the SRS microscope to the fingerprint region, or is ~2800–3050 cm^-1 the full range for your setup? 2850 and 2920 cm^-1 appear sufficient to support your findings, but I'm curious if there are wavenumbers in the fingerprint region you would have wanted to image at?

The SRS images of the worms are stunning! I agree that this dataset would be a valuable resource, but has it been made public? I did not see a statement regarding data availability. Especially because you have correlated 2PF with the SRS images, it would be a rich and unique dataset to share with the community!

Nice study! I am wondering though if you could you provide an explanation of how the sensitivity for these measurements is "comparable" for both 532nm and 633nm excitations when only 633nm has a resonance effect?

It appears that the data point for the absorbance measurement at the highest MFU is missing in (C) -- I think because it is covered by the data point from the Raman data? This nicely shows that the two methods have the same qualitative results, but quantifying the correlation would also be good.

Based on the name of the instrument (RL637), I think the earlier mentions of 633nm should instead be 637nm, as stated here and throughout the rest of the manuscript.

Additionally, if there are two measurable peaks when exciting with 633nm vs. only one measurable peak for 532nm, is there any advantage to exciting with 532nm?

The results you obtain with SpotMAX are impressive compared to those from human annotation and what you've identified as the SOTA solution. I'm left wondering, though, where a method less sophisticated than the SOTA but more automated and convenient than manual annotation (such as e.g., trackpy.locate) would fall in this comparison.

Apologies if I've overlooked it, but I couldn't find more information about how this is done in the Methods section. From my experience, inferring this information from image acquisition parameters is relatively straightforward for any microscope you have data from (because you can find the relevant properties in the metadata), but it's very difficult in general (because every manufacturer outputs metadata in their own way). I'd be interested to know if you have a general solution for this problem using e.g. Bio-Formats for Python.

I understand the logic, but how can you be sure that a plateau in the increase in fluorescence lifetime means that the microtubules have completely depolymerized?

Could you please provide a link to the GitHub repository? Is it publicly available yet?

This is a very cool visualization! The choice of the colormap is, however, not the most intuitive, especially given that the same color is used to represent different concentrations in different panels.

What is meant by "identity" masks? I understand these masks to be based on either the lifetime or the number of photons per pixel, what does identity mean in this context?

This seems like a great case study for showcasing your software! But if you are going to make this claim, could you explain how or why reference [35] is not the first? Given that you also say:

Changes in the fluorescence lifetime of SiR-tubulin following the addition of Nocodazole have been used to quantify the degree of microtubule depolymerisation [35].

While I don't think it has been published yet, there is also PhasorPy, which is an open-source, Python-based library for doing phasor analysis.

Fully agree! Would it be possible to add references to alternative approaches? I am also curious if you considered revising the segmentation based on unmatched nuclei. One could argue that the isolated nucleus in Fig. 2 should be re-labeled as not a nucleus.

As required by both models? Would it be possible to comment briefly on to what extent InstanSeg is scale-invariant?

Is the RGB projection truly random or are the colors evenly spaced?

Is the ground truth training data balanced between nucleus and cell labels or is this not necessary?

I'm curious if your light microscopy data used for cell tracking is, or could also be, made available?

Cool video! I'm curious though if any filtering of the trajectories was done to prevent individual cells from being counted multiple times in the statistics? Relatively faster cells that dip in and out of focus, for instance, could skew the speed distribution if not taken into consideration.

Sorry, but this also confuses me. How can this be considered a global transformation when only one tile is matched at a time for translation? I would have assumed that the tiles are roughly in place from using e.g. stage coordinates as a crude estimate of tile position.

Why do you say the algorithm can be used to "only find the scaling and rotation parameters" when you immediately add translation? Do you mean only scaling and rotation are solved for in the original implementation of the algorithm and your adaptation adds translation?

Your geometric hashing approach sounds similar to RANSAC. I am curious if you compared the two and what the tradeoffs between time and accuracy look like?

Hello, I am not very familiar with this imaging modality but it seems rather strange that the imaging parameters are unknown. Is this typically the case?

This is indeed an impressive field of view, and the apparent SNR is impressive for ~10s exposures. More details on the laser power, diffuser, and detection optics would be appreciated!

Could you annotate the reference line in the plot to make it more clear where the pixel intensities in (c) are extracted from.

Could you provide information on the laser you used for this? It would be great if you could provide a more complete description of your experimental setup in a Methods section.

It's surprising to me that dark current poses more of a problem than autofluorescence. I'm curious if the dark current is somehow exacerbated when detecting in the SWIR region or if it would be mitigated simply with better detector cooling?

Intrinsically denoised, sure, but noisy training data may still lead to a smoothed/blurred prediction. Are there techniques you have experimented with to sharpen predictions or has this not been an issue in your experience?

For many applications the virtual staining could be seen as a means to an end: segmentation. Is the intermediate step of predicting the virtual stain necessary or would it be feasible to skip straight to the segmentation by e.g. merging the Cellpose segmentation model?

not to mention dataset sizes.

Can this claim be quantified without deferring to a figure?

The class representations of cell type 1 and cell type K are not very well differentiated.

IMO Supplemental figures 1 and 2 would be more clear and succinct as heatmaps than line plots.

You provide a clear explanation of how you generate the pseudospectra in the methods, but I question the use of "ideal" in describing them. I get that you mean idealized, but it sounds as if they are perfect and in no way subjective which is misleading.

Could you please clarify if this test was done by using manual segmentation as ground truth? If not, what was the source of ground truth data used to compute F1-scores?

An F1-score of 0.7 is not generally seen as very high. Could emphasize more strongly in your discussion the power of your approach (combining the instance segmentation with class maps for classification) given that this seems to be a non-trivial segmentation problem.

Is this claim quantitatively verified?

How is this a fair comparison for comparing SNR between live and cryofixed conditions? I understand your point that exposure time for live Raman is limited due to photodamage, but would still be interesting to see 30s/line for live Raman (albeit with drift artifacts) for the sake of comparison.

*exposure, compensating

This is a nice example of benchmarks for band widths for certain spectral peaks to differentiate between low and high SNR!

Can you more clearly define what constitutes "high-SNR" Raman? Of course it is high relative to "low-SNR" Raman, but how can someone know if they are doing high- vs low-SNR Raman? Are there thresholds that are (or can be) defined on the width of certain spectral peaks?

Is fused silica necessary to prevent autofluorescence from surface defects in normal glass coverslips? Is it typical/recommended to used fused silica for Raman microscopy generally or only for very high sensitivity measurements?

*show

*wavelength dispersion

Would be helpful to cite a couple specific examples of biological activities and/or molecular species for which fixation is insufficient to bolster motivation for cryogenic freezing.