A powerful new multi-view deconvolution algorithm [with video]

Modern light-sheet fluorescence microscopy is capable of imaging the development of an entire specimen at cellular level with high temporal resolution and provides unique opportunities for the observation of the same specimen from multiple angles (views). Deconvolution, a computational post-processing technique, can significantly increase spatial resolution of such datasets, but processing times are prohibitively long for large acquisitions. We developed a new multi-view deconvolution algorithm that takes into account conditional dependencies between views and provides a fast implementation that exploits the computing power of graphics cards. In combination, these improvements can achieve an up to 120-fold faster processing of light-sheet microscopy datasets.

HFSP Long-Term Fellow Stephan Preibisch and HFSP Young Investigator Grant holder Pavel Tomancak and colleagues
authored on Mon, 19 May 2014

Biologists are never satisfied with the images coming from their microscopes. They always want them sharper, showing more details and less noise. One way to achieve that is, of course, to improve the microscopy itself. Alternatively, a computational technique called deconvolution has been used to estimate the underlying real image given the imperfectly acquired image and knowledge about the optical properties of the microscope. The results of the deconvolution can be impressive, however, it is notoriously difficult to calculate, especially for large three-dimensional images acquired over long periods of time.

Stephan Preibisch previously in the group of Pavel Tomancak, and now an HFSP fellow with Robert Singer and Eugene Myers, took a fresh look on the deconvolution technique using data from microscopes that allow observation of the biological specimen from multiple angles. This multi-view microscopy makes the difficult problem of guessing the underlying image from the microscopic observations more tractable. Preibisch showed that by exploiting the multiple observations of the same specimen and their dependency, it is possible to speed up the deconvolution process several fold. Together with Fernando Amat from HHMI Janelia Farm Campus, he developed algorithms and software that exploit the processing power of the graphics card to perform multi-view deconvolution in minutes rather than hours. The results are stunning. For example, on a multi-view acquisition of C. elegans larva deconvolution easily distinguishes all the organism’s nuclei from one another. Preibisch applied the algorithm to data from several different model organisms, including recordings of Drosophila embryonic development captured over many hours by a Selective Plane Illumination Microscope (SPIM). His approach is the only one capable of dealing with such large datasets in reasonable time.

Figure 1: Worm. Comparison of input data and deconvolved image for fixed Caenorhabditis elegans larva in L1 stage expressing LMN-1–GFP (green) and stained with Hoechst (magenta). The top two rows show input data of orthogonal views acquired by the Zeiss Lightsheet Z.1 microscope, the last row shows the result of the multi-view deconvolution.

The algorithms can be applied to different microscopes since the software is open source and can be adapted to different imaging modalities. It is distributed as a plugin for the Fiji platform where it extends previously developed algorithms for multi-view registration. The complete suite of plugins is applicable to multi-view SPIM data generated by the commercially available Lightsheet Z.1 from Carl Zeiss Microimaging as well as ‘do it yourself’ OpenSPIM set-ups.

Figure 2: Drosophila OpenSPIM1. Comparison of image quality for an OpenSPIM acquisition of a Drosophila embryo expressing His-YFP in all cells. The top row shows the quality of one of the six input views along all three orientations. The center row illustrates the image after content-based fusion, while the last row shows the same image after multi-view deconvolution.

Reference

Efficient Bayesian-based Multiview Deconvolution. Stephan Preibisch, Fernando Amat, Evangelia Stamataki, Mihail Sarov, Robert H. Singer, Eugene Myers, Pavel Tomancak (2014). Nature Methods, doi:10.1038/nmeth.2929.

Other References

[1] OpenSPIM - an open access platform for light sheet microscopy. Pitrone P., Schindelin J., Stuyvenberg L., Preibisch S., Weber M., Eliceiri K.W., Huisken J., Tomancak P. (2013). Nature Methods, 10, 598–599.

Video 1: Developing Drosophila embryo I

Video 2: Developing Drosophila embryo II

Video 3: C. elegans multi-view deconvolution I

Video 4: C. elegans multi-view deconvolution II

PubMed link

Nature Methods article

Preprint at Arxiv

Fiji wiki

Article from Pavel Tomancak on OpenSpim