GaussMLE.jl

Fast Maximum Likelihood Estimation of Gaussian PSF parameters for single-molecule localization microscopy. Automatic GPU acceleration with CPU fallback.

Installation

using Pkg
Pkg.add("GaussMLE")

Quick Start

using GaussMLE

# Input: ROIBatch from your detection pipeline
batch = ROIBatch(data, x_corners, y_corners, frames, camera)

# Fit: returns (BasicSMLD, GaussMLEFitInfo)
smld, info = fit(batch, GaussMLEConfig(
    psf_model    = GaussianXYNB(0.13f0),  # fixed-σ Gaussian, σ in μm
    backend      = :auto,                  # :cpu, :gpu, or :auto (GPU w/ CPU fallback)
    iterations   = 20,                     # Newton-Raphson iterations
    batch_size   = 10_000,                 # ROIs per GPU kernel launch
))

# Output: positions and CRLB uncertainties in μm
smld.emitters[1].x, smld.emitters[1].σ_x

For complete SMLM workflows (detection + fitting + rendering), see SMLMAnalysis.jl.

GPU scheduling uses NVML polling to avoid OOM during device discovery in multi-process environments. See the GPU guide for details.

PSF Models

All PSF parameters are specified in microns for camera-independence.

Model	Constructor	Fitted Parameters	Use Case
GaussianXYNB	GaussianXYNB(σ)	x, y, N, bg	Fixed-width 2D Gaussian
GaussianXYNBS	GaussianXYNBS()	x, y, N, bg, σ	Variable-width 2D Gaussian
GaussianXYNBSXSY	GaussianXYNBSXSY()	x, y, N, bg, σx, σy	Anisotropic 2D Gaussian
AstigmaticXYZNB	AstigmaticXYZNB{T}(σx₀, σy₀, Ax, Ay, Bx, By, γ, d)	x, y, z, N, bg	3D astigmatic localization

Output Format

fit() returns (BasicSMLD, GaussMLEFitInfo). The emitter type in BasicSMLD depends on the PSF model:

PSF Model	Emitter Type	Extra Fields
GaussianXYNB	Emitter2DFitGaussMLE	pvalue
GaussianXYNBS	Emitter2DFitSigma	σ, σ_σ, pvalue
GaussianXYNBSXSY	Emitter2DFitSigmaXY	σx, σy, σ_σx, σ_σy, pvalue
AstigmaticXYZNB	Emitter3DFitGaussMLE	z, σ_z, σ_xz, σ_yz, pvalue

All emitters include: x, y, photons, bg, σ_x, σ_y, σ_xy, σ_photons, σ_bg, pvalue, frame.

GaussMLEFitInfo fields:

Field	Description
elapsed_s	Wall time (seconds)
backend	:cpu or :gpu (never :auto)
device_id	GPU index (0-based) or -1 for CPU
n_fits	Number of ROIs processed
n_converged	Number converged (currently = n_fits)
batch_size	Actual batch size used
n_batches	Number of batches
memory_per_batch	Estimated bytes per batch

Performance

Benchmarked on AMD Ryzen Threadripper PRO 5975WX / NVIDIA RTX A6000, 11x11 ROIs, 10K fits:

Model	CPU (fits/s)	GPU (fits/s)
GaussianXYNB (ideal)	2,120	401,572
GaussianXYNB (sCMOS)	2,077	261,590
GaussianXYNBS (ideal)	1,633	215,745
AstigmaticXYZNB (ideal)	1,341	265,083

GPU throughput scales with batch size; these numbers use the default batch size of 10K. Run Pkg.test("GaussMLE") locally to benchmark your hardware.

Algorithm Reference

MLE fitting algorithm:

Smith, C., Joseph, N., Rieger, B. et al. "Fast, single-molecule localization that achieves theoretically minimum uncertainty." Nat Methods 7, 373-375 (2010). DOI: 10.1038/nmeth.1449

sCMOS pixel-dependent noise model:

Huang, F., Hartwich, T.M.P., Rivera-Molina, F.E. et al. "Video-rate nanoscopy using sCMOS camera-specific single-molecule localization algorithms." Nat Methods 10, 653-658 (2013). DOI: 10.1038/nmeth.2488

Related Packages

• SMLMAnalysis.jl - Complete SMLM workflow (detection + fitting + rendering)
• SMLMData.jl - Core data types for SMLM
• SMLMSim.jl - SMLM data simulation
• MicroscopePSFs.jl - PSF models

License

MIT License - see LICENSE file for details.