This module allows you to train and use deep learning based models for the denoising of images. Microscope image quality can be compromised by various imperfections originating from electronic and thermal sources, which introduce noise into images. This noise affects image quality across different microscopy modalities, including fluorescence and electron microscopy, complicating the differentiation between genuine signal and noise.

Fluorescence microscopy images, especially those of live cells, often suffer from a low signal-to-noise ratio due to limitations on excitation energy to prevent phototoxicity or other damage to cells. Techniques such as averaging the signal over multiple frames or slowing the scanning speed can improve the signal-to-noise ratio, but may still result in noisy images. Electron microscopy images are susceptible to noise from factors such as low electron dose, specimen drift, and detector noise. Non-conductive materials must be imaged at ultra-low voltages and currents to prevent beam charging effects and sample damage, resulting in noisy images. To remove noise, traditional denosing algorithms, such as Gaussian filtering, non-local means filtering, etc., can be used, which typically come with trade-offs like reduced image sharpness. Deep Learning-based algorithms, including Noise2Void, offer more robust and efficient denoising solutions.

Intellesis Denoising allows you to train and use deep learning based models for the denoising of images. It supports any type of images (except for RGB) and any dimensions and it is not dedicated to a special field of application. Intellesis Denoising allows the training of both 2D and 3D denoising models. The 3D denoising models are especially suitable for achieving more consistent denoising results in 3D datasets. Intellesis Denoising uses Noise2Void (N2V) and Noise2Void2 (N2V2), which require only a noisy input image for the training of a model and can thus be trained directly on the data that should be denoised.

To give a simplistic explanation, N2V replaces pixels by masked pixel data randomly selected within a certain window/surrounding area. With this approach, the model is then trained to reconstruct the original pixels and to discard the implicit noise in the image. For detailed information on N2V, refer to the paper "Noise2Void - Learning Denoising from Single Noisy Images" written by Alexander Krull, Tim-Oliver Buchhol and Florian Jug, see https://arxiv.org/abs/1811.10980. Intellesis Denoising also enables you to use Noise2Void2 (N2V2), which is a modification of Noise2Void to reduce unwanted checkerboard artifacts.

It is important to ensure that the trained denoising model is compatible with the images to be denoised in terms of pixel size and bit depth, noise characteristics, structural features, and imaging modality for optimal results. AI denoising should be used before applying processing functions that modify the pixel values to that image. Note that Intellesis Denoising requires the installation of Docker Desktop and the respective denoising container, which can both be installed with the ZEISS Microscopy Installer.

A short acquisition time (0.2 ms) results in a noisy raw image (left image), but the signal-to-noise ratio (SNR) can be improved significantly by AI Denoising post-processing (center image). A high-SNR image (acquisition time 150 ms) is shown as reference (right image).