M-Denoiser: Unsupervised image denoising for real-world optical and electron microscopy data.

Area of Science:

• Microscopy imaging
• Computer vision
• Image processing

Background:

• Real-world microscopy images suffer from significant noise, including signal-dependent shot noise and signal-independent read noise, often modeled by Poisson-Gaussian distributions.
• Noise in microscopy data is frequently spatially correlated due to the image acquisition process.
• Existing unsupervised and self-supervised denoising methods often assume signal-independent or pixel-wise independent noise, which is unsuitable for real-world microscopy data.

Purpose of the Study:

• To develop an unsupervised denoising method, M-Denoiser, specifically designed for real-world microscopy data.
• To overcome the limitations of current denoising algorithms that fail to account for signal-dependent and spatially correlated noise.

Main Methods:

• The proposed M-Denoiser utilizes a 'shatter module' to decorrelate noise before the denoising process.
• A novel unsupervised training loss function is introduced, specifically designed for pairs of noisy microscopy images.
• The model was trained and evaluated on both optical and electron microscopy datasets.

Main Results:

• M-Denoiser demonstrated superior performance in denoising real-world microscopy images compared to existing baseline methods.
• Quantitative and qualitative evaluations confirmed the effectiveness of the proposed method.
• The shatter module and novel loss function contribute to improved denoising accuracy.

Conclusions:

• M-Denoiser offers an effective unsupervised solution for denoising real-world microscopy images.
• The method successfully addresses the challenges posed by signal-dependent and spatially correlated noise.
• This advancement has the potential to improve the quality and utility of microscopy data in scientific research.