Loading…
Fluorescence microscopy denoising for transient synaptic signals, done frame by frame so one checkpoint transfers across sensors and frame rates.
No providers recorded yet. Browse all providers
A glutamate transient at a single synapse is about the hardest thing a fluorescence microscope is asked to report: the release site is orders of magnitude smaller than a cell body, the sensor rises and decays within a couple of frames, and each frame collects very few photons. Denoisers built for calcium imaging borrow signal from neighbouring time points, interpolating a frame from those around it or splitting a stack into interlaced sub-stacks. That works for slow somatic calcium, blurs fast synaptic glutamate, and welds the trained network to one acquisition rate and one sensor's kinetics.
Neuroimage Denoiser removes the temporal dimension from the problem entirely. A 2D U-Net denoises each frame on its own, trained by the Noise2Noise principle: extra Gaussian noise is added to an already-noisy crop and the network is asked to recover the original noisy frame, so no noise-free ground truth is ever needed. Because nothing in the model sees frame ordering, the same weights apply whether the camera ran at 100 Hz or 1000 Hz and whether the reporter is a fast glutamate sensor or a slower calcium one.
The framework comes from Martin Heine's and Susanne Gerber's groups at the Institute of Developmental Biology and Neurobiology and the Institute of Human Genetics of Johannes Gutenberg University Mainz and University Medical Center Mainz, released as a bioRxiv preprint. Code, a single pretrained checkpoint, and the training data are all public.
The network is a U-Net with one input and one output channel. Encoder blocks pair two 3×3 convolutions with ReLU and batch normalization followed by 2×2 max-pooling, widening from 64 through 128 and 256 to 512 feature maps; the decoder mirrors this with transposed convolutions and skip concatenation, ending in a 1×1 convolution with a linear output. The paper and repository state no parameter count; the 124 MB PyTorch checkpoint implies roughly 31 million parameters at fp32, consistent with a 64-to-512 U-Net of this shape. With no dense layers, inference accepts frames of any size, bounded only by GPU memory. Training used 32×32 z-transformed crops selected at a minimum z-score of 3, batch size 64, the ADAM optimizer at a learning rate of 0.0001, a smooth L1 loss, and Gaussian noise with center 0 and scale 2.0, the combination chosen by a grid search ranking amplitude error against residual noise.
The released checkpoint saw only iGluSnFR3.v857.SGZ recordings acquired at 100 Hz from DIV16–19 primary mouse hippocampal cultures on an inverted TIRF setup with a scientific CMOS camera. Applied unchanged, it improved peak signal-to-noise ratio 5.65 ± 0.33-fold at 100 Hz and 5.21-, 7.72-, and 8.97-fold at 200, 500, and 1000 Hz, with amplitude correlation R = 0.9991 across those rates. On sensors absent from training it gave a 2.00 ± 0.12-fold SNR gain for iGluSnFR.S72A (R = 0.9998) and 4.47 ± 0.23-fold for synaptophysin-GCaMP6f (R = 0.9995). In one fast pulse train, evoked responses the raw trace missed became detectable, cutting the apparent failure rate from 40% to 0%.
The immediate audience is laboratories doing camera-based functional imaging at single-synapse resolution, where noise obscures the quantities of interest: release probability, failure rate, paired-pulse ratio, and the separation of adjacent release sites within a complex synapse. Distribution is a pip-installable command-line tool that reads TIFF and Nikon ND2 stacks, denoises whole directories while preserving folder structure, and writes TIFFs that drop into existing Fiji or Python trace-extraction workflows. Groups with different optics or sensors can retrain on their own recordings using the same preparation.
Neuroimage Denoiser makes a deliberate trade — discarding the temporal context other denoisers depend on — and shows that for spatially confined, temporally sparse signals spatial structure alone suffices to restore the frame while leaving amplitudes intact, so one checkpoint covers many acquisition rates and sensors rather than one condition each. The evidence base is correspondingly narrow: training drew on a single sensor at one frame rate from one optical setup, and every transfer result comes from the same laboratory's recordings, with no external cohort, in vivo, or scanning-microscope validation. The authors scope the claim to sensors "when the localization is similar", and the weakest transfer was to the lower-amplitude iGluSnFR.S72A. Code and weights carry a non-commercial, no-derivatives Creative Commons license (CC BY-NC-ND 4.0), whose terms conflict with the CC BY-NC 4.0 badge in the README — under the license file as written, a fork or a fine-tuned checkpoint could not be redistributed. The work remains a preprint.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.