Source-linked AI summary
A Poisson-Gaussian Denoising Dataset with Real Fluorescence Microscopy Images
Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, Scott Howard
TL;DR
Fluorescence microscopy lacks a reliable real-image dataset dedicated to its predominantly Poisson-Gaussian noise. This paper constructs and evaluates the FMD dataset using averaged ground truths and multiple noise levels, finding that deep-learning methods perform best and can support fast denoising.
Problem
Reliable datasets of real fluorescence microscopy images for Poisson-Gaussian denoising were lacking, despite the need to reduce noise in weak-signal microscopy.
Method
The paper constructs the FMD dataset from real confocal, two-photon, and wide-field images, estimates ground truths by averaging captures, creates five noise levels, and benchmarks 10 denoising algorithms.
Results
Deep-learning models outperform the other eight benchmarked methods across noise levels in PSNR and SSIM, while Noise2Noise achieves similar performance to DnCNN without clean images.
Takeaways & Limitations
The publicly available FMD dataset provides a benchmark for Poisson-Gaussian fluorescence microscopy denoising and may support improved real-time imaging quality.
Takeaways & Limitations
Image registration is not performed because the commercial microscope image sequences are assumed to be already well aligned.
Abstract
from arXiv · showhide
Fluorescence microscopy has enabled a dramatic development in modern biology. Due to its inherently weak signal, fluorescence microscopy is not only much noisier than photography, but also presented with Poisson-Gaussian noise where Poisson noise, or shot noise, is the dominating noise source. To get clean fluorescence microscopy images, it is highly desirable to have effective denoising algorithms and datasets that are specifically designed to denoise fluorescence microscopy images. While such algorithms exist, no such datasets are available. In this paper, we fill this gap by constructing a dataset - the Fluorescence Microscopy Denoising (FMD) dataset - that is dedicated to Poisson-Gaussian denoising. The dataset consists of 12,000 real fluorescence microscopy images obtained with commercial confocal, two-photon, and wide-field microscopes and representative biological samples such as cells, zebrafish, and mouse brain tissues. We use image averaging to effectively obtain ground truth images and 60,000 noisy images with different noise levels. We use this dataset to benchmark 10 representative denoising algorithms and find that deep learning methods have the best performance. To our knowledge, this is the first real microscopy image dataset for Poisson-Gaussian denoising purposes and it could be an important tool for high-quality, real-time denoising applications in biomedical research.
1. Introduction
Fluorescence microscopy provides important biomedical imaging capabilities but produces inherently weak, predominantly Poisson-noisy signals, creating a need for effective denoising and reliable evaluation data. The FMD dataset addresses this gap with real images, averaged ground truths, and multiple noise levels.
- Fluorescence microscopy is widely used in biomedical research, including confocal, two-photon, and wide-field modalities.
- Ground truth images in the dataset are estimated by averaging 50 noisy raw images from fluorescence microscopy acquisitions.Figure 1 illustrates single-channel mouse-brain images, multi-channel BPAE-cell images, different noise levels, and their averaged ground truths.
- Weak photon signals make fluorescence microscopy images inherently noisy and dominated by Poisson rather than Gaussian noise.The detector captures approximately 10^2 photons per pixel, compared with approximately 10^5 in photography.
- Increasing excitation power cannot indefinitely improve image cleanliness because biological light dosage and fluorescence saturation constrain it.
- Rapid imaging may preclude longer acquisition times because each image can need to be captured within tens of milliseconds.
- The FMD dataset contains 12,000 real noisy images spanning three microscopy modalities and representative samples including cells, zebrafish, and mouse brain tissues.Image averaging produces ground truth images and noisy images at five different noise levels.
2. Related Work
Existing real-image denoising datasets primarily target Gaussian or general Poisson-Gaussian noise rather than Poisson-dominated fluorescence microscopy. The paper characterizes FMD noise and benchmarks traditional and deep-learning approaches on real noisy microscopy images.
- Real-image denoising datasets such as RENOIR, Darmstadt Noise, Smartphone Image Denoising, and PolyU lack a dedicated focus on Poisson-dominated microscopy images.The paper distinguishes fluorescence microscopy noise from Gaussian-denoising datasets because shot noise is dominant.
- Image averaging is the most commonly used method for obtaining ground truth images when constructing denoising datasets.
- FMD avoids image registration because commercial microscope sequences are already well aligned, with estimated global translations below half a pixel.The authors state that sub-pixel translation would smooth noisy images and undermine the realness of Poisson noise.
- Poisson-Gaussian denoising can use methods designed for the noise model directly, such as PURE-LET, or variance-stabilizing transformations followed by Gaussian denoisers.The transformation-based route supports methods including NLM, BM3D, KSVD, EPLL, and WNNM.
- Deep-learning denoisers include supervised models such as DnCNN and blind or unpaired approaches such as FFDNet, CBDNet, UDNet, and Noise2Noise.Noise2Noise does not require clean images and previously outperformed VST+BM3D on synthetic Poisson noise by almost 2 dB.
- FMD noise is Poisson-dominated for two-photon and confocal microscopy but has a larger Gaussian component in wide-field microscopy.Benchmarking 10 representative algorithms shows better real-image denoising performance for deep-learning models than for traditional methods.
3. Noise Modeling in Fluorescence Microscopy
Fluorescence microscopy measurements are modeled with signal-dependent Poisson noise and signal-independent Gaussian noise. The section formulates denoising and explains variance stabilization and averaging for producing cleaner estimates.
- Noise model: The Poisson-Gaussian model combines signal-dependent shot noise with signal-independent, zero-mean Gaussian noise.The detector scaling coefficient is a, while b denotes Gaussian-noise variance.
- Denoising problem: The denoising problem estimates the underlying ground-truth pixel yi from noisy measurement zi.
- Variance stabilization: Variance-stabilizing transformation converts Poisson-Gaussian denoising into approximately unit-variance Gaussian denoising.Gaussian methods such as NLM and BM3D can then be applied, followed by an inverse transformation.
- Variance stabilization: The closed-form inverse VST approximation preserves denoising accuracy while reducing computational cost.The resulting estimate is the noise-free signal eyi.
- Image averaging: Averaging independent noisy images keeps the mean equal to the ground truth while reducing estimation noise and increasing SNR.The dataset uses S = 1, 2, 4, 8, 16 for noise levels and S = 50 for ground truth; little improvement appears beyond about 40 images.
4. Dataset
The FMD dataset uses repeated low-excitation acquisitions across microscopy modalities and biological samples to create aligned noisy images, averaged ground truths, and multiple noise levels. It contains 240 FOVs, 12,000 raw images, and 60,000 noisy realizations, with modality-dependent noise characteristics.
- 4.1. Image Acquisition Setup: FMD covers confocal, two-photon, and wide-field microscopy using commercial microscopes and real biological samples.
- 4.2. Noisy Image and Ground Truth Estimation: Registration found translations below 0.5 pixel, so the commercially captured sequences were already well aligned and registration was omitted.Sub-pixel translation could smooth noisy images and damage the real Poisson-noise characteristics.
- 4.2. Noisy Image and Ground Truth Estimation: Averaging S = 2, 4, 8, 16 raw images creates four additional noise levels, while all 50 captures provide each FOV’s ground truth.The shared ground truth preserves the same underlying signal across noise realizations.
- 4.1. Image Acquisition Setup: Each of 12 imaging configurations contains 20 FOVs, with 50 repeated raw captures per FOV.
- 4.3. Dataset Statistics and Noise Estimation: 240 FOVs yield 12,000 raw images and 60,000 noisy images after five-fold extension by circular averaging.
- 4.3. Dataset Statistics and Noise Estimation: Confocal and two-photon images are Poisson-dominated, whereas wide-field images have a larger Gaussian component.This conclusion comes from grouped estimates of the noise parameters a and b.
5. Benchmark
The benchmark evaluates ten denoising algorithms on real Poisson-Gaussian fluorescence microscopy images across mixed modalities and noise levels. Deep learning models generally outperform traditional methods, while Noise2Noise approaches DnCNN without clean training images and can achieve very fast GPU inference.
- 5.1. Setup: Ten representative algorithms are benchmarked on mixed fluorescence microscopy test sets spanning confocal, two-photon, and wide-field images.The mixed test set uses five noise levels and averages metrics over 48 noise realizations per level.
- 5.2. Results and Discussion: Deep learning models outperform the other eight methods by a significant margin in PSNR and SSIM across all noise levels.The models are blind to noise levels and are retrained on the FMD dataset.
- 5.2. Results and Discussion: Noise2Noise has similar denoising performance to DnCNN without requiring clean images and achieves almost 2 dB higher PSNR than VST+BM3D.Noise2Noise performs slightly better than DnCNN in the high-noise domain.
- 5.2. Results and Discussion: Batch normalization stabilizes Noise2Noise training at a relatively large learning rate but does not improve PSNR when the learning rate is well tuned.The comparison uses learning rates of 0.001 and 0.0001.
- 5.2. Results and Discussion: Residual learning improves DnCNN denoising, whereas training Noise2Noise sufficiently long allows it to outperform DnCNN and VST-BM3D.Figure 5 reports this behavior after, for example, 400 epochs.
- 5.2. Results and Discussion: GPU inference reduces deep learning denoising time to less than 1 ms, potentially enabling real-time processing at up to 100 frames per second.CPU denoising times are similar to those of VST-BM3D and PURE-LET.
6. Conclusion
The paper concludes by presenting the FMD dataset as a dedicated benchmark for real fluorescence microscopy images with Poisson-Gaussian noise. Deep learning models trained on it outperform other methods across imaging modalities and noise levels.
- 6. Conclusion: The FMD dataset contains real fluorescence microscopy images with Poisson-Gaussian noise and covers most microscopy modalities.Image averaging provides ground truth and noisy images at five noise levels.
- 6. Conclusion: The benchmark includes raw single-channel zebrafish embryo images and raw multi-channel BPAE cell images evaluated with PSNR and SSIM.Color-image values are averaged across the three channels.
- 6. Conclusion: Deep learning denoising models trained on FMD outperform other methods by a large margin across all imaging modalities and noise levels.The dataset is publicly available as a benchmark for Poisson-Gaussian denoising research.
A.1. Fluorescence microscopy setup
The dataset uses commercial confocal, two-photon, and wide-field fluorescence microscopy setups with specified objectives, image sizes, pixel sizes, acquisition settings, and preprocessing.
- A.1. Fluorescence microscopy setup: Confocal and two-photon images were acquired on a Nikon A1R-MP microscope with a 40×, 1.15 NA water-immersion objective.These images were 512×512 pixels with 300 nm pixels and a 2 µs pixel dwell time.
- A.1. Fluorescence microscopy setup: Wide-field images were acquired on a Nikon Eclipse 90i microscope with a 40×, 0.75 NA objective and a camera exposure time of 200 ms.The 1280×960 raw images were cropped to 512×512 before dataset processing.
A.2. Pixel clipping or over/under-exposure
The microscopy setup was adjusted to limit detector saturation and pixel clipping, while the dataset does not identify under-exposure as a practical issue.
- A.2. Pixel clipping or over/under-exposure: Manual detector-gain adjustment limited clipped pixels to at most 0.2% of pixels in all images.Clipping can occur when bright biological structures saturate the detector.
- A.2. Pixel clipping or over/under-exposure: Under-exposure is not considered an issue because fluorescence microscopy detectors have high sensitivity and accuracy.The primary exposure concern described is over-exposure from bright fluorescence signals.
A.3. Benchmark results on separate test set
The separate-test-set benchmark evaluates denoising across imaging configurations and noise levels using repeated noise realizations. Deep learning methods almost dominate traditional methods in every test case.
- 50 noise realizations are evaluated for each imaging configuration and noise level on the pre-selected 19th FOV test set.Results are organized in Tables 6–8 for confocal, two-photon, and wide-field microscopy.
- The benchmark covers confocal, two-photon, and wide-field microscopy configurations.
- Deep learning-based denoising methods almost dominate traditional methods across all test cases.
A.4. Network architecture and training details
The evaluated networks retain the original DnCNN and Noise2Noise structures, with a modified N2N-BN variant. Training uses standardized 256 × 256 inputs, shared optimization settings, 400 epochs, and minibatches formed from cropped patches.
- Network architecture: DnCNN and Noise2Noise largely preserve their original network structures, while N2N-BN adds batch normalization and a final Tanh activation.
- Training details: Inputs are normalized to [−0.5, 0.5], and Adam optimization uses β1 = 0.9, β2 = 0.99, and zero weight decay.
- Training details: Training lasts 400 epochs with one-cycle learning-rate scheduling, rising to 0.0001 before cosine annealing.
- Training details: Both DnCNN and Noise2Noise use minibatches of 16 constructed from four 512 × 512 training images cropped into non-overlapping 256 × 256 patches.
- Evaluation: Confocal, two-photon, and wide-field denoising results report PSNR, SSIM, and denoising time averaged over 50 noise realizations.