⚠ JavaScript is disabled in your browser. webSMLM runs entirely in JavaScript on your own machine (nothing is uploaded), so it cannot start without it. Please enable JavaScript for this page and reload.

webSMLM v0.11.0 · proof-of-concept

Memory & streaming

Memory budget (GB) Stream heap (MB)

Simulation settings

Frames Pixel size (nm) Emitter density (µm⁻²·frame⁻¹) Photons/emitter/frame ON lifetime (frames, mean) Gain (photons/ADU) Camera offset (ADU) Offset std (ADU, per-pixel) Read noise σ (e⁻) Background (photons/px) Drift (px, total)

Gain & offset estimation

Frames sampled k_thresh (spatial freq.) Readout noise σ (e⁻)

module: fit

Estimates camera gain (photons/ADU) and offset (ADU) directly from the loaded stack via the Rieger–Heintzman photon-conversion-factor method (PCFO; Heintzmann, Relich, Nieuwenhuizen, Lidke & Rieger, arXiv:1611.05654) — no separate calibration acquisition needed. Tiles a sample of frames, measures mean signal vs. high-spatial-frequency (noise-only) variance per tile, and fits gain/offset by linear regression, and draws a diagnostic signal-vs-noise-variance scatter + fitted line (R²) on the raw panel so the underlying linearity assumption can be checked visually rather than trusted blindly. Estimate only computes — it doesn't touch the Gain / Camera offset fields in Localisation settings itself; Transfer estimates (disabled until a successful Estimate) copies the last result into them.

Frames sampled — how many seeded-random frames to average over.
k_thresh — spatial-frequency cutoff (fraction of Nyquist) above which content is assumed noise-only; 0.9 follows the Rieger–Heintzman default.
Readout noise σ — camera dark-frame std, in electrons; set to 0 if unknown/negligible. Must include ANY spatially-white, frame-invariant noise, not just true per-frame read noise — a static per-pixel offset pattern (fixed-pattern noise) looks identical to read noise in a single frame's Fourier content, so it biases the fitted offset the same way if left out. If both are present, combine them in quadrature (√(read_noise²+offset_std²), in photon-equivalent units); the default (2.89) matches the Simulation panel's own default read noise (2.7 e⁻) and per-pixel offset std (3 ADU) combined this way, for a clean self-test.

Tile size isn't a setting — it's chosen automatically from the field of view (aiming for a ~4×4 grid, rounded to a power of two since the noise-variance estimate needs an FFT, floored at 64px, and never below 2×2) so there's nothing to re-tune per stack.

3D calibration

First frame Last frame z-step (nm) z=0 ref frame (0=auto) Fix bead x,y (robust at high defocus)

Localisation settings

Real-time update Fit method Temporal median filtering First frame Last frame Detection filter Exact band-pass (slower, DoG only) Threshold (k·σ_noise) Threshold (k·σ_noise) Threshold (intensity) σ_PSF (px) Fit radius (px)

module: detect/fit

Fit method — σ_noise below is the spread of the filtered image, not the PSF width.

Phasor — ~100–250× faster per candidate (265× on the GATTA-PAINT test stack). No per-localization uncertainty.
Gaussian LS — ordinary least-squares; ≈ slightly better precision than phasor, much slower.
Gaussian MLE (default) — Poisson-optimal, reports a real CRLB uncertainty; cost ≈ LS.

Detection filter — Wavelet and DoG both band-pass the frame (suppress smooth background, enhance PSF-sized spots); candidates are strict local maxima above k·σ_noise. The three filters respond very differently, so re-tune the threshold when switching between them.

Wavelet (B-spline) (default) — the à trous cubic-B-spline wavelet used by ThunderSTORM: no σ (scale is fixed by the wavelet levels), roughly 2× faster to filter, and the recommended choice.
DoG band-pass — a difference of Gaussians whose scale is tuned by σ_PSF; its Exact band-pass option replaces the fast box approximation of the background with a true Gaussian (~2× slower, but exactly reproducible — on the GATTA-PAINT test stack it changes 0.38% of detections).
Uniform box filter — a difference of two box (uniform) averages sized off σ_PSF, following Huang, Schwartz, Byars & Lidke (2011); unlike the other two it thresholds on a plain intensity value, not k·σ_noise, so its default (25) needs re-tuning to your camera's counts.

Temporal median filtering (FTM) is a per-pixel background correction — for a given frame, each pixel's value across a window of nearby frames (centred on that frame; clamped, not shrunk, at the very first/last few frames, so every frame still gets a full-width window) has its median subtracted — a robust estimate of that pixel's slowly-varying background at this point in time. Since a blinking emitter occupies far less than half the window at any one pixel, the median tracks the background, not the signal, so what's left after subtraction is mostly signal above background. Checking this box:

Adds a raw / FTM-corrected toggle next to the raw panel title, computing the correction live for whichever frame you're scrubbed to.
Makes Localize itself run on FTM-corrected frames — processed in worker-parallel chunks when a worker pool is available, main-thread otherwise; either way the whole stack is never held twice (raw + corrected) in memory at once.

Values that would go negative after subtraction are floored at the camera offset instead (a background estimate briefly above the true noise floor is expected, not an error) — this keeps the corrected pixel in the same units the fit's own gain/offset conversion expects, so photon counts aren't double-corrected. It's still not quite the same noise distribution the fit functions otherwise assume (Poisson-MLE in particular models a Poisson-distributed background around some positive mean; a hard floor truncates that slightly, which can bias fitted background/photon counts and reported uncertainty low in very dim regions). Choose the window size so the "signal occupies less than half the window" assumption above actually holds for your data (too small and the median starts tracking — and cancelling — real signal); the loaded stack must have at least that many frames.

The technique originates with Nieuwenhuizen et al., Nat. Methods 10, 557–562 (2013); ported here from the Hohlbein Lab's own newer implementation, FTM2 (Jabermoradi et al., Phil. Trans. R. Soc. A 380(2220), 20200164, 2022).

Pixel size (nm) Gain (photons/ADU) Camera offset (ADU)

Rendering settings

Magnification Render blur σ_render (px) Colour map Display max

Drift correction (AIM)

Segment (frames) Search radius (nm)

Localization precision (NeNA / FRC)

Spectral SMLM analysis

Distance min (nm) Distance max (nm) Primary angle (deg) Angle tolerance (± deg) Require narrower 0th order (σ)

module: sSMLM

Pairs 0th/1st-order localizations from a diffraction grating placed in the emission path — each emitter appears twice per frame, offset by a wavelength-dependent distance at a fixed, known bearing (not just orientation — Primary angle is a genuine direction, e.g. 0° always means the 1st order sits to the same side of every 0th order in the image). A point qualifies as a 0th order only if it has a candidate on that bearing AND no candidate on the opposite bearing (which would mean it's more likely someone else's 1st order) — this needs no brightness signal, since real data shows brightness alone doesn't reliably tell 0th from 1st order here. The paired position is the 0th order's own — undispersed, so its centroid is the true emitter position — not the midpoint between the two (that would blur position by up to half the per-emitter spectral offset). The inter-order distance is stored in z, so the existing depth-coding render option (Rendering settings → Colour by depth) shows it directly as a wavelength proxy with no other change needed. Localizations that don't find an unambiguous pair within the window are dropped from the result entirely.

Preview pairs only computes — Show dist. hist. draws a distance histogram (every candidate pair in range, any angle); Show angle hist. draws an angle histogram restricted to the current distance window, so you can find your own setup's true peak instead of guessing. Both histograms are accumulated across ALL frames (only same-frame localizations are ever compared to each other — the accumulation just pools every frame's own candidates into one plot). Narrow Distance min/max and Primary angle/tolerance to that peak, then click Pair to commit — or click Fit angle & tol. to fill Primary angle/Angle tolerance in automatically from the angle histogram's peak (its half-max width), a conservative starting point you can widen by hand.

Pair replaces the current localizations with one row per accepted pair (refuses if the current result already has real 3D z from an astigmatic fit method, since sSMLM pairing and 3D depth both use the same field). Unpair restores the original, unpaired localizations.

Require narrower 0th order (σ) is an optional extra confidence gate: the 0th order is undispersed while the 1st is spectrally smeared, so it tends to have the narrower PSF — but only ~65–70% reliably on real data, so this is off by default rather than required.

2-point pairs only (0th+1st) for now — multi-order chaining is not yet implemented, see docs/REFACTOR_PLAN.md. Ported from HohlbeinLab/sSMLMAnalyzer — see References & further reading in Help & guide.

Raw frame — detected spots overlaid in green

SMLM reconstruction — run localization to reconstruct

Log

webSMLM v0.11.0 Ready. Click “Load movie” or “Simulate movie”, then “Localize”.

Localizations

— in the raw panel, uses the current filter

…more info…

[module: table]

Filtering — field op value, joined with and/or (e.g. intensity > 1000 and uncertainty < 20); Enter commits a clause, clauses stack cumulatively. Type reset to clear all of them. Start typing a field name for autocomplete suggestions (↑/↓ to move, Enter/Tab to accept) — covers both the current table's columns and the tempClusteringXY/tempClusteringZ fields below.
sigma_xy / sigma_z — sigma_xy is the fitted PSF width (average of σx/σy for astigmatic 3D). sigma_z, shown for MLE 3D results, is an approximate z-precision: the fit's own uncertainty on σx/σy propagated through the calibration curve's local slope at the found z — not a full joint fit (z isn't a parameter of the pixel-level fit, σx/σy are), but the best estimate available from what the fit reports. Not available for Phasor 3D.
Temporal clustering — tempClusteringXY < 10 (nm) merges the same molecule's detections across consecutive frames into one higher-precision "event" (photon-weighted position, summed intensity, properly combined uncertainty); tempClusteringZ < value adds the same in z. One threshold per axis (a new value replaces the old). Once active, counts switch from localizations to events and the table gains an nmerged column. A gap frame always breaks the chain for now (planned: tempClusteringMemory). FRC resolution can get coarser after clustering, not finer — counter-intuitive, but FRC measures achieved image resolution, which depends on point density as well as per-point precision; merging improves the latter but sharply reduces the former (far fewer points), and if that density loss outweighs the precision gain, FRC's number gets worse even though individual events really are more precise. Check the uncertainty column's histogram (not FRC) to see the precision improvement in isolation.
Plot histogram of — draws the chosen column's distribution in the raw (left) panel, over whatever rows currently pass the filters above; the crop tool and typed filters affect it the same way.

webSMLM — how it works

A proof-of-concept single-molecule localization pipeline that runs entirely in your browser. Nothing is uploaded.

Guided workflow

Get data. Click Load movie for a real .tif/.tiff stack, or Simulate movie for a test dataset (open Simulation settings to adjust frames, blink density and photons).
Set detection & fit. Pick a Fit method first — the relevant parameters follow it. Adjust Threshold (k·σ_noise) if too many/few spots are boxed in the raw view, and set σ_PSF to match your spot size.
Localize. The raw view refreshes live with detected ROIs (green boxes) and accepted localizations as the run progresses; the right panel builds the super-resolution image.
Explore the result. Set Pixel size (nm) (e.g. 100 or 160) — it defines the physical scale for the scale bar and export. Change Magnification/σ_render to re-render instantly (no refit). Scroll or pinch to zoom, drag to pan, double-click/tap to reset.

1 · Loading & memory

TIFF headers are parsed first to get the frame count and dimensions, then the decoded working-set size is estimated (frames × width × height × 4 bytes).

If it fits the memory budget (default 3 GB, adjustable) all frames are decoded into RAM with a progress bar — changing the threshold and re-running is then instant (no re-decode).

If it exceeds the budget, the loader streams: frames are decoded as the analysis reaches them and discarded again, so memory stays bounded no matter how long the stack is. Trade-off: re-running re-decodes. For ordinary TIFFs this happens in heaps of ~500 MB (adjustable) because every frame must pass through the TIFF decoder; one heap is decoded, localized, freed, then the next. This is exactly the pattern that would let you localize a live camera buffer in real time.

Large ImageJ stacks (above ~4 GB) are written with a single directory entry and all frames laid out contiguously after it, because 32-bit TIFF offsets cannot address further. webSMLM detects this from the images=N record and indexes frames arithmetically, decoding one at a time — no heaps needed.

The file itself is never read into memory whole — frames are pulled on demand via File.slice(), so a multi-GB stack stays bounded by the memory/chunk budget above regardless of on-disk size.

Bit depth (8/16/32-bit) and endianness are handled automatically; 16-bit is preserved (not flattened to 8-bit).

2 · Detection — finding ROIs

Band-pass filter (Difference of Gaussians). Each frame is blurred at a small scale (0.8×PSF σ) and a large scale (3×PSF σ); the large-scale blur is subtracted from the small-scale one. This removes flat background and slow gradients while keeping PSF-sized spots. DoG = G(0.8σ) − G(3σ).

Local-maximum picking with a k·σ threshold. On the filtered image we compute the mean and standard deviation of pixel values, and keep any strict 8-neighbour local maximum whose value exceeds mean + k·σ. Here σ is the noise spread of the band-passed image (not the PSF width), and k is the slider: k = 4 keeps peaks > 4 noise-standard-deviations above background. Lower k → more (dimmer) detections; higher k → fewer, more confident ones. The Wavelet (B-spline) filter (default) replaces the DoG band-pass with an à trous cubic-B-spline wavelet plane, but is thresholded the same k·σ way.

Uniform box filter. Instead of Gaussians, the band-pass is a difference of two box (uniform) averages sized off σ_PSF, and candidates are strict local maxima of that band-pass — over a σ_PSF-sized square window, not just the 8 neighbours — above a plain intensity threshold rather than k·σ. It has no per-frame noise statistic, so the default threshold (25) is in raw camera counts and needs re-tuning per dataset.

Each surviving maximum defines a small ROI (Fit window, e.g. 7×7 px) passed to the fitter.

3 · Fitting — sub-pixel position

Phasor (fast). Takes the phase of the first Fourier coefficient of the ROI in x and y to get the sub-pixel centre. No iteration — ~100–170× faster than Gaussian fitting, ~19 nm precision at 900 photons. Great for large stacks and real-time.

Gaussian (least-squares). Fits a 2-D Gaussian to the ROI by Gauss–Newton with a line search (~16 nm at 900 photons). Slightly better precision, much slower.

Gaussian (MLE). Poisson maximum-likelihood fit of an integrated-Gaussian PSF (Smith et al. 2010), by full Fisher-scoring Newton (~5 iterations, cost similar to LS). Statistically optimal at low photon counts, and it reports a proper per-localization CRLB uncertainty (the “uncertainty [nm]” column comes from the Cramér–Rao bound rather than a formula estimate). Candidates whose fit didn't converge within its iteration budget, whose intensity collapsed to the amplitude floor (background mistaken for a spot), or whose CRLB came out non-finite are rejected automatically rather than kept as degenerate localizations.

Note: Gaussian (LS) is ordinary least-squares; Gaussian (MLE) is the Poisson-optimal fit — prefer MLE at low photon counts. Phasor is the speed option.

4 · Rendering

Localizations are splatted into a high-resolution accumulation image at the chosen Magnification (super-res pixels per camera pixel) and optionally smoothed by Render blur. Both re-render from stored localizations without refitting. The scale bar is computed from your pixel size and the current zoom; scroll/drag/double-click to navigate.

Precision figures above are for the synthetic model; real data depends on your photon count, background and PSF.

References & further reading

What this tool borrows from, and where to read more:

Phasor localization

“Phasor based single-molecule localization microscopy in 3D (pSMLM-3D): an algorithm for MHz localization rates using standard CPUs,” K. J. A. Martens, A. N. Bader, S. Baas, B. Rieger, J. Hohlbein, J. Chem. Phys. 148, 123311 (2018). doi:10.1063/1.5005899
“Integrating engineered point spread functions into the phasor-based SMLM framework,” K. J. A. Martens et al., Methods (2020).

Spot detection & thresholding

“ThunderSTORM: a comprehensive ImageJ plug-in for PALM and STORM data analysis and super-resolution imaging,” M. Ovesný, P. Křížek, J. Borkovec, Z. Švindrych, G. M. Hagen, Bioinformatics 30(16), 2389–2390 (2014). doi:10.1093/bioinformatics/btu202
“Wavelet analysis for single molecule localization microscopy,” I. Izeddin et al., Opt. Express 20(3), 2081–2095 (2012). doi:10.1364/OE.20.002081
“Simultaneous multiple-emitter fitting for single molecule super-resolution imaging,” F. Huang, S. L. Schwartz, J. M. Byars, K. A. Lidke, Biomed. Opt. Express 2(5), 1377–1393 (2011). doi:10.1364/BOE.2.001377

Fitting & precision (why LS ≠ MLE)

“Precise nanometer localization analysis for individual fluorescent probes,” R. E. Thompson, D. R. Larson, W. W. Webb, Biophys. J. 82(5), 2775–2783 (2002). doi:10.1016/S0006-3495(02)75618-X
“Optimized localization analysis for single-molecule tracking and super-resolution microscopy,” K. I. Mortensen, L. S. Churchman, J. A. Spudich, H. Flyvbjerg, Nat. Methods 7, 377–381 (2010). doi:10.1038/nmeth.1447
“Fast, single-molecule localization that achieves theoretically minimum uncertainty,” C. S. Smith, N. Joseph, B. Rieger, K. A. Lidke, Nat. Methods 7, 373–375 (2010). doi:10.1038/nmeth.1449
MLE implementation ported from Picasso's picasso/gaussmle.py — see Picasso reference below.

Gain/offset estimation (PCFO)

“Calibrating photon counts from a single image,” R. Heintzmann, P. K. Relich, R. P. J. Nieuwenhuizen, K. A. Lidke, B. Rieger, arXiv:1611.05654. arxiv.org/abs/1611.05654

Localization precision (NeNA)

“A simple method to estimate the average localization precision of a single-molecule localization microscopy experiment,” U. Endesfelder, S. Malkusch, F. Fricke, M. Heilemann, Histochem. Cell Biol. 141, 629–638 (2014). doi:10.1007/s00418-014-1192-3

Image resolution (FRC)

“Measuring image resolution in optical nanoscopy,” R. P. J. Nieuwenhuizen, K. A. Lidke, M. Bates, D. L. Puig, D. Grünwald, S. Stallinga, B. Rieger, Nat. Methods 10, 557–562 (2013). doi:10.1038/nmeth.2448

Temporal median filtering (FTM)

Originates with the Nieuwenhuizen et al. paper above; ported from the Hohlbein Lab's own newer implementation, FTM2, used in “Enabling single-molecule localization microscopy in turbid food emulsions,” A. Jabermoradi, S. Yang, M. I. Gobes, J. P. M. van Duynhoven, J. Hohlbein, Phil. Trans. R. Soc. A 380(2220), 20200164 (2022). doi:10.1098/rsta.2020.0164

Drift correction (AIM)

“Toward drift-free high-throughput nanoscopy through adaptive intersection maximization,” H. Ma, M. Chen, P. Nguyen, Y. Liu, Sci. Adv. 10(21), eadm7765 (2024). doi:10.1126/sciadv.adm7765
Adapted from Picasso's picasso/aim.py (parabolic sub-pixel peak fit replaces the FFT phase refinement; linear interpolation replaces the spline) — see Picasso reference below.

Picasso (reference implementation for the ported MLE and AIM drift code above, github.com/jungmannlab/picasso)

“Super-resolution microscopy with DNA-PAINT,” J. Schnitzbauer, M. T. Strauss, T. Schlichthaerle, F. Schueder, R. Jungmann, Nat. Protoc. 12, 1198–1228 (2017). doi:10.1038/nprot.2017.024

Overview

“Single-molecule localization microscopy,” M. Lelek et al., Nat. Rev. Methods Primers 1, 39 (2021). doi:10.1038/s43586-021-00038-x

Acknowledgements

Contributions beyond the author:

Koen J. A. Martens contributed the uniform box-filter detection method (Huang et al. 2011, see above), the Real-time update live-preview toggle, the wheel-zoom / ctrl+wheel-scrub raw-panel scroll rework, and sizing the green ROI box from the actual fit radius (PR #4, v0.9.1).
Koen J. A. Martens also contributed Gain & offset estimation (the PCFO photon-conversion-factor method, see above) and the Simulation settings module's physically-motivated emitter-lifetime and camera forward models (PR #5, v0.10.2).

License & author

webSMLM © 2026 Hohlbein et al., licensed under CC BY 4.0. You may share and adapt it, including commercially, with attribution.
Bundled decoders retain their own licenses: UTIF.js and pako are MIT-licensed.