A Poisson-Gaussian Denoising Dataset with Real Fluorescence Microscopy Images (1812.10366v2)

Abstract: 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.

• The paper introduces the Fluorescence Microscopy Denoising (FMD) dataset, comprised of 12,000 real microscopy images with synthesized ground truths, to address Poisson-Gaussian noise prevalent in this imaging modality.
• Evaluation of 10 diverse denoising algorithms on the FMD dataset shows deep learning techniques significantly outperform traditional methods in image quality metrics like PSNR and SSIM.
• The FMD dataset provides a robust benchmark for algorithm development and validation, with implications for improving image quality and enabling faster analysis in fluorescence-based biological research.

Overview of "A Poisson-Gaussian Denoising Dataset with Real Fluorescence Microscopy Images"

The paper addresses a notable deficiency in the field of fluorescence microscopy image processing by introducing the Fluorescence Microscopy Denoising (FMD) dataset. This dataset is meticulously crafted to assist in the denoising of Poisson-Gaussian noise prevalent in real microscopy images. Key distinctions of fluorescence microscopy images include noise levels far superior to typical photography due to the weak photon signals, posing challenges for image denoising and necessitating specialized algorithms and datasets.

The newly curated FMD dataset is comprised of 12,000 real, noisy microscopy images captured using commercial confocal, two-photon, and wide-field microscopes, exemplifying biological samples such as BPAE cells, mouse brain tissues, and zebrafish embryos. The ground truth images are synthesized through the technique of image averaging—fifty repetitions of the same field of view—to mitigate noise effects and establish baseline clarity absent in individual noisy captures. The dataset further augments its value by encompassing images at varying noise levels, corresponding to different numbers of contributing averaged captures, thereby producing a rich testing ground for denoising algorithms.

In the paper, the authors rigorously evaluate 10 representative denoising algorithms spanning classical approaches and emerging deep learning models. Noteworthy among these are methods like VST+BM3D, DnCNN, and Noise2Noise. Of particular interest is the finding that deep learning techniques surpass traditional methods significantly in performance metrics such as PSNR and SSIM, enhancing overall image quality even without explicit clean image datasets. Additionally, the authors detail the computational time required for various models, noting that deep learning approaches, while initially demanding in training, demonstrate exceptional speed during inference—opening avenues for their application in real-time biomedical imaging situations. Such capabilities may substantially enhance dynamic and time-sensitive biological research by providing clearer visualizations sooner than previously achievable.

An essential aspect of this work is the detail in noise modeling specific to fluorescence microscopy imaging systems, which the researchers illustrate using a Poisson-Gaussian noise model. This model accommodates the signal-dependent Poisson component and an additive Gaussian component to better reflect the characteristics of optical signals within this imaging domain. The paper also discusses transformation techniques such as variance-stabilizing transformation (VST) that enable tackling Poisson-Gaussian denoising with established Gaussian noise removal methodologies.

Implications of the FMD dataset are broadly felt. Practically, the dataset offers a robust, reliable foundation for evaluating and comparing denoising algorithms, potentially advancing the quality and veracity of results in fluorescence-based biological research. Theoretically, the dataset advocates and enables the further refinement and diversification of noise reduction models, particularly deep learning genera capable of handling and learning from the nuanced characteristics under low signal conditions encountered in microscopy.

Looking ahead, the establishment of such a dataset may spur advances in image processing technologies specifically tailored for low-light and photon-limited scenarios. The burgeoning success of deep learning models in this context, as expounded by this paper, suggests fruitful research directions in unsupervised and semi-supervised learning paradigms, potentially exploring synergistic applications in diverse imaging systems beyond the biological scope. Consequently, the paper's contributions dwell not just in addressing current deficiencies in fluorescence imaging but also in guiding the trajectory of future developments within the discipline.