Knowing when to stop: a data-driven termination criterion for iterative deconvolution

Session

Reproducibility of Data Analysis at Scale

Authors

Jan Becker (4), Craig Russell (2), Andrew York (1), James Manton (3)

Affiliations

1. Calico Life Sciences LLC
2. EMBL-EBI
3. MRC Laboratory of Molecular Biology
4. University of Oxford

Keywords

deconvolution, image processing, iterative algorithm, stopping criterion

Abstract text

Iterative deconvolution is a commonly used and powerful tool to improve contrast and apparent resolution in light microscopy datasets. However, there is a tricky problem in determining the correct number of iterations to apply: too few and the deconvolved data will be overly blurry; too many and the deconvolved data will contain artificial small-scale structures derived from over-approximating the noise in the original image. As the correct number of iterations depends on the signal-to-noise ratio of the image data and the structure of interest, it is not possible to determine ahead of time, limiting the ability to include high-quality deconvolution in a processing pipeline.

While a number of regularised deconvolution algorithms have been proposed, these all have the drawback that manual tweaking of the iteration number is merely replaced by manual tweaking of the regularisation coefficients. Here, we present an alternative approach that determines an optimal iteration number purely from the original image data. We show how an optimal number of iterations can be derived from the noise statistics of the image and apply our method with Richardson-Lucy deconvolution. Furthermore, we demonstrate that our stopping criterion can be changed from a global stopping criterion to a local one, allowing high-SNR regions of the image to be deconvolved to a sharper result while leaving low-SNR regions artifact free. We show results from automatically processed 2D and 3D datasets from a variety of microscopy modalities and explain how similar results can be obtained using our open-source code.