Improving electron micrograph signal-to-noise with an atrous convolutional encoder-decoder.

Area of Science:

• Microscopy
• Image Processing
• Deep Learning

Background:

• Electron microscopy generates high-resolution images crucial for scientific research.
• Image noise, particularly in low-dose conditions, can obscure fine details and hinder analysis.
• Existing denoising methods often struggle to preserve structural integrity while effectively reducing noise.

Purpose of the Study:

• To develop and evaluate a deep learning-based denoising method for electron micrographs.
• To improve the signal-to-noise ratio in electron microscopy images without sacrificing resolution.
• To provide a robust and efficient tool for enhancing the quality of electron microscopy data.

Main Methods:

• An atrous convolutional encoder-decoder neural network architecture was employed.
• The network features a modified Xception backbone, atrous convolutional spatial pyramid pooling, and a multi-stage decoder.
• Training involved simulated low-dose data derived from high-dose micrographs with added Poisson noise, followed by fine-tuning on actual high-dose data.

Main Results:

• The proposed neural network significantly outperformed traditional methods (bilateral, Gaussian, median, total variation, wavelet, Wiener) in denoising electron micrographs.
• Performance gains were measured by mean squared error (MSE) and structural similarity index (SSIM), achieving 24.6% and 9.6% improvements for low doses, and 43.7% and 5.5% for high doses, respectively.
• The neural network demonstrated the lowest variance in mean squared error across both low- and high-dose datasets.

Conclusions:

• The developed atrous convolutional encoder-decoder network is highly effective for denoising electron micrographs.
• This deep learning approach offers superior performance compared to conventional denoising techniques.
• The availability of source code, datasets, and pre-trained models facilitates further research and application in electron microscopy.