Denoising: AI versus Huygens Deconvolution

Efficient de-noising and de-blurring without the need for training data

Comparing AI de-noising with Huygens true deconvolution

A major challenge in microscopy is to obtain images with a high signal to noise ratio (SNR). Various de-noising methods have been developed to improve the SNR which can be categorized into Noise statistics, AI/deep-learning, and Deconvolution. Unlike the first two methodologies, which basically ignore the presence of out-of focus signal (blur) due to the point spread function (PSF), true deconvolution has the strong advantage that it combines the correction of noise, background and the PSF. When comparing deconvolution with other (de-noising) methods, it is important to realize that not all deconvolution algorithms and their implementations perform equally well^1,2,3,4,5, or have been used optimally⁶.
Here, we compared our Huygens deconvolution with CARE (Content-Aware Reconstruction)⁷, one of the well-known Deep Learning de-noising methods^8,9. Also, we show an extreme low SNR confocal example (SNR=2) and how well Huygens can restore objects from such noisy data. A download link is included to let you reproduce this result (see below).

Conclusions: Based on our results (outlined below and available upon request), we conclude that Huygens deconvolution significantly outperforms the AI de-noising method CARE. Advantages of Huygens deconvolution:

Much faster as it works directly on the dataset

No training data needed

Better noise correction without risk of creating hallucinations

Better signal increase and background correction (see figure below)

Restoration of out-of focus signal/PSF (see figure below)

Improved 3D resolution both latteraly and axially.

Top: Both CARE AI restoration and Huygens clearly increase the SNR. However, CARE produces hallucinations: it misidentifies background noise as object signal in raw image data (arrowheads), a known response of AI ^10,11. By contrast, Huygens deconvolution does not.

Bottom: CARE AI generates objects from imaging artifacts (like a hot pixel), whereas Huygens accounts for and corrects these distortions properly.

Huygens Deconvolution leads to better image restoration compared to CARE

$DiffractionLimit CARE$

Both CARE AI denoising and Huygens Deconvolution correct for noise, but only deconvolution is capable of properly correcting the background and out-of-focus signal, which are clearly visible as PSFs in the image. Note the intensity peak in the graph, indicating the significant object-specific gain in signal and resolution achieved with Huygens deconvolution

Huygens Deconvolution can restore objects from extremely low noise image data

Here is a challenge! Try reproducing this Huygens Deconvolution result yourself. This noisy raw 3D confocal dataset of yeast cells has a calculated SNR of 2, with intensity values as shown in the left histogram (the maximum count is 5 photons). It took the fully automated Huygens Deconvolution Express less than 20 seconds to deconvolve this 3D data. Huygens succesfully restored objects, the existence of which can be confirmed using other images (with a different noise distribution of the same sample). Note the significant increase in signal and dynamic range in the histograms before (left) and after Huygens deconvolution (right), and conversion to 16bit unsigned integer *.tif. Use the button below to download the raw confocal 3D dataset as an ics2 file, which can be opened via Bioformats. Feel free to reproduce the result also with Huygens.

Download the dataset Free Huygens Trial

1xp1 Ch0 Ch0 Rawhistogram 1xp1 Ch0 Decon 2 Ch0 Histogram

Raw (left) and Huygens deconvolved (right) results of a noisy confocal 3D image. The image was acquired with a Leica SP8 at Nyquist sampling 40x40x120 nm (xyz) using photon counting mode. The raw image was opened in Huygens and deconvolved with standard settings using Deconvolution Express (under the Deconvolution menu. Top shows a single slice, and bottom shows a 3D MIP animation. Image courtesy: Prof. Benjamin Glick and Kasey Day, Dept of MCDB, University of Chicago, USA

CARE compared to Huygens Deconvolution

References

1. van Kempen, GMP., van der Voort, HTM., Bauman, JGJ., & Strasters, KC. (1996). Comparing maximum likelihood estimation and constrained Tikhonov-Miller restoration. IEEE Engineering in Medicine and Biology Magazine, 15(1), 76-83.
2. van Kempen, GMP, van Vliet LJ, Verveer P.J, van der Voort, HTM (1997) A quantitative comparison of image restoration methods for confocal microscopy. Journal of Microscopy, Vol. 185, Pt 3, 354–365.
3. Liu, Y., Panezai, S., Wang, Y. et al. Noise amplification and ill-convergence of Richardson-Lucy deconvolution. Nat Commun 16, 911 (2025). https://doi.org/10.1038/s41467-025-56241-x
4. Li, Y., Su, Y., Guo, M. et al. Incorporating the image formation process into deep learning improves network performance. Nat Methods 19, 1427–1437 (2022). https://doi.org/10.1038/s41592-022-01652-7%%% 5. https://svi.nl/MaximumLikelihoodEstimation
6. https://forum.image.sc/t/deconvolution-vs-ai-deblurring-for-microscopy/84896/2
7. Weigert M, Schmidt U, Boothe T, Müller A, Dibrov A, Jain A, et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. Nature Methods. 2018 Nov 26;15(12):1090–7. Available from: https://www.nature.com/articles/s41592-018-0216-7%%% 8. Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Karras, T., Aittala, M. and Aila, T. Noise2Noise: Learning image restoration without clean data. arXiv. 2019.
9. Krull, A., Buchholz, T.O. and Jug, F. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.
10. https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)
11. Leonhard Möckl, Anish R. Roy, and W. E. Moerner, "Deep learning in single-molecule microscopy: fundamentals, caveats, and recent developments (Invited)," Biomed. Opt. Express 11, 1633-1661 (2020) https://doi.org/10.1364/BOE.386361

Scientific Volume Imaging

Deconvolution - Visualization - Analysis