FM2S: Towards Spatially-Correlated Noise Modeling in Zero-Shot Fluorescence Microscopy Image Denoising

Abstract

Fluorescence microscopy image (FMI) denoising faces critical challenges due to the compound mixed Poisson-Gaussian noise with strong spatial correlation and the impracticality of acquiring paired noisy/clean data in dynamic biomedical scenarios. While supervised methods trained on synthetic noise (e.g., Gaussian/Poisson) suffer from out-of-distribution generalization issues, existing self-supervised approaches degrade under real FMI noise due to oversimplified noise assumptions and computationally intensive deep architectures. In this paper, we propose Fluorescence Micrograph to Self (FM2S), a zero-shot denoiser that achieves efficient FMI denoising through three key innovations: 1) A noise injection module that ensures training data sufficiency through adaptive Poisson-Gaussian synthesis while preserving spatial correlation and global statistics of FMI noise for robust model generalization; 2) A two-stage progressive learning strategy that first recovers structural priors via pre-denoised targets then refines high-frequency details through noise distribution alignment; 3) An ultra-lightweight network (3.5k parameters) enabling rapid convergence with 270× faster training and inference than SOTAs. Extensive experiments across FMI datasets demonstrate FM2S's superiority: It outperforms CVF-SID by 1.4dB PSNR on average while requiring 0.1% parameters of AP-BSN. Notably, FM2S maintains stable performance across varying noise levels, proving its practicality for microscopy platforms with diverse sensor characteristics. Code and datasets will be released.

Visualization

Getting Started

It is recommended to use a virtual environment.

conda create -n FM2S python=3.9 -y
conda activate FM2S
pip install -r requirements.txt

Data

Our experiment data are from the following links, and we acknowledge their work!

• FMD (Zhang et al. 2019) dataset. yinhaoz/denoising-fluorescence

• SRDTrans (Li et al. 2023) dataset. cabooster/SRDTrans

Two subset for our experiment are in the data folder, one can use them to reproduce our results.

Run

The starting program is in main.py and run the following command to denoise an image.

python main.py -i data/TwoPhoton_MICE_1.png -o output.png -c twophoton

The arguments are explained as:

• -i/--input_image_path: The path for image to be denoised.
• -o/--output_image_path: The output path for the denoised image.
• c/--config: The microscope type of the input image. All the configurations are listed below.

FMD dataset

Type	HypParam	avg1	avg2	avg4	avg8	avg16
Confocal	$\lambda$			2
	stride			75
	filter			3
	$k_g$	200	125	70	10	5
	$k_p$	30	95	195	240	650
	$\lambda_p$	70	285	485	650	1400
TwoPhoton	$\lambda$			2
	stride			75
	filter			3
	$k_g$	175	150	90	20	15
	$k_p$	30	85	300	185	850
	$\lambda_p$	60	300	480	600	3800
WideField	$\lambda$	1
	stride			75
	filter			11
	$k_g$	220	220	60	20	1
	$k_p$	45	100	650	600	1500
	$\lambda_p$	2000	2500	3500	4000	4800

SRDTrans dataset

$\lambda$	stride	filter	$k_g$	$k_p$	$\lambda_p$
6	5	3	60	40	125

Citation

@article{liu2026fm2s,
  title={FM2S: Towards Spatially-correlated Noise Modeling in Zero-shot Fluorescence Microscopy Image Denoising},
  author={Liu, Jizhihui and Teng, Qixun and Ma, Qing and Hou, Junhui and Jiang, Junjun},
  journal={Machine Intelligence Research},
  volume={23},
  number={1},
  pages={200--213},
  year={2026},
  publisher={Springer}
}