An interactive ImageJ plugin for semi-automated image denoising in electron microscopy

Three-dimensional electron microscopy (3D EM) techniques, such as serial block face scanning electron microscopy (SBF-SEM) and focused ion beam scanning electron microscopy (FIB-SEM), provide high-resolution 3D images of biological samples. However, the acquisition of large-scale datasets can take days or even months, especially at high resolutions. To reduce acquisition time, shorter dwell times are often employed, resulting in increased noise levels in the acquired images. Advanced denoising techniques can mitigate this problem, but many existing methods are challenging to use due to their reliance on low-level programming environments, complex parameter tuning, or high computational demands. This necessitates the development of user-friendly tools that combine state-of-the-art denoising algorithms with an intuitive interface, allowing for efficient parameter optimization and rapid processing of large datasets. The current study addresses this need by introducing DenoisEM, a novel plugin designed to enhance the accessibility and efficiency of image denoising in 3D EM.

Existing solutions for EM image restoration are often impractical for large-scale 3D datasets due to their reliance on low-level programming languages, complex parameter tuning, and high computational costs. While many denoising methods exist (e.g., multiresolution shrinkage, nonlocal pixel averaging, Bayesian estimation, convolutional neural networks), they lack the user-friendliness required for broad adoption. Interactive frameworks often suffer from difficult-to-interpret parameters or excessive computational demands. The lack of a readily accessible and extensible framework that incorporates state-of-the-art algorithms highlights the need for an improved solution that balances computational efficiency with ease of use and expert interaction.

DenoisEM is an ImageJ plugin designed to provide an interactive and GPU-accelerated workflow for denoising 3D EM images. The plugin guides the user through a six-step process: data loading, initialization, region-of-interest (ROI) selection, noise estimation, interactive parameter optimization, and final batch processing. The user selects a representative ROI, and the plugin automatically estimates the noise level to provide near-optimal parameter initialization. The interactive interface allows users to adjust parameters and view the results in real time, facilitated by the GPU-accelerated backend using the Quasar framework. Eight denoising algorithms are implemented: Gaussian filtering, wavelet thresholding, anisotropic diffusion, bilateral filtering, Tikhonov denoising/deconvolution, total variation denoising, Bayesian least-squares Gaussian scale mixtures (BLS-GSM), and nonlocal means denoising/deconvolution. The plugin also includes a blur estimation feature. The workflow allows for easy switching between algorithms and parameters without overwriting the original image. The plugin saves algorithm parameters as metadata, ensuring reproducibility. Specific details on the implementation of each algorithm, including noise estimation, blur estimation, and parameter estimation (using polynomial fitting), are provided in the methods section of the paper.

DenoisEM significantly accelerates the denoising process compared to existing software, achieving speed improvements of one to two orders of magnitude for various algorithms. The plugin allows for a fourfold increase in data acquisition speed by enabling the use of shorter dwell times without sacrificing image quality. Denoising with DenoisEM improves visualization and interpretation of ultrastructure in various EM datasets. In experiments with *Arabidopsis thaliana* root tip and mouse heart tissue SBF-SEM images, denoising facilitated better recognition of subcellular structures like the nuclear membrane and endoplasmic reticulum, and improved the distinction of sarcomeres. Denoising also enhances automated segmentation. In mouse heart tissue FIB-SEM data, denoising improved the accuracy of automated sarcomere counting, reducing discrepancies with manual counts. In experiments with the CREMI dataset, denoising improved the quality of automated neuronal membrane segmentation, even at high noise levels, suggesting that acquisition times can be significantly reduced without affecting segmentation performance. The GPU acceleration makes parameter tuning efficient, allowing for near real-time visualization of denoising effects.

DenoisEM addresses the critical need for faster and more accessible denoising in 3D EM. The plugin's user-friendly interactive interface and GPU acceleration overcome limitations of existing methods. The fourfold increase in acquisition speed demonstrated is particularly significant, as it translates to substantial time savings in large-scale imaging projects. The results confirm that denoising with DenoisEM improves image quality, visualization, and the accuracy of automated segmentation analyses. This suggests that researchers can obtain comparable results with much shorter acquisition times, enabling higher throughput and reducing the demands on imaging resources. The broader accessibility of sophisticated denoising techniques through DenoisEM may also facilitate research in areas previously limited by computational constraints.

DenoisEM is a valuable contribution to the field of electron microscopy, providing an accessible and efficient tool for image denoising. The plugin's speed, user-friendliness, and the inclusion of multiple state-of-the-art algorithms significantly improve the workflow for processing large 3D EM datasets. Future work could focus on predictive parameter optimization based on machine learning, extending the plugin to handle multichannel data, and incorporating newer denoising techniques such as those based on deep learning.

The current version of DenoisEM focuses on grayscale images. While the algorithms themselves are not inherently limited to 2D, the current implementation processes volumes slice-by-slice for computational efficiency. Although the plugin offers a wide range of algorithms, the optimal parameters for each algorithm and dataset may still require expert adjustment. The user-friendliness of the plugin is relative and assumes a basic level of familiarity with ImageJ.