microscope is a powerful tool for characterizing the structure and dynamics of cells and subcellular . While advances in microscopy technology provide great promise for quantitative and precise measurement of morphological and molecular phenomena at the single-cell level of bacteria, quantitative analysis of microscopic images remains a challenge, as many bacteria can only be visible in the range of visible light wavelengths.
Two years ago, scientists developed a deep learning-based segmentation method called Cellpose, which accurately segments cells from various image types without the need for model retraining or parameter adjustment. However, this method is still limited by methods independent of cell morphology or optical properties.
Recently, in a new study published on "Nature Methods" , a research team from University of Washington developed the deep neural network image segmentation algorithm - Omnipose. Omnipose trained with a large bacterial image database performs well in describing and quantifying various bacteria in mixed microbial cultures and eliminates possible errors when using its predecessor tool Cellpose.
As we all know, bacteria has a variety of forms in addition to its tiny figure. Although the bacteria commonly studied are basically similar to rod-shaped or spherical, some bacteria are in other forms, such as spiral. In addition, the microfluidic device allows researchers to capture bacteria’s response to various treatments such as antibiotic , but in this process it can often lead to irregular morphology of bacteria. These morphological properties make it difficult for existing deep learning tools to determine which bacteria exist from microscopic images.
In this new study, the team developed the deep neural network image segmentation algorithm Omnipose. Unique network outputs (such as gradients of distance fields) allow Omnipose to accurately segment existing algorithms (including its predecessor, Cellpose) to produce errors.
study shows that Omnipose achieves unprecedented segmentation performance on mixed bacterial cultures, antibiotic-treated cells, and cells in elongated or branched morphology. In addition, Omnipose does a better job of overcoming the identification problems caused by differences in optical properties of different bacteria. The algorithm-enhanced cell segmentation performance promotes unique insights into microbial systems.
Omnipose can accurately identify poisoned E. coli cells
However, there are some limitations when reproducing crowded microbial community 3D samples in 2D. For example, overlapping elements in an image can create an illusion effect. Furthermore, when the bacterial size is below a certain threshold, Omnipose is difficult to detect.
Nevertheless, when analyzing cells from the raw dataset of fast-growing Arabidopsis roots, Omnipose can show some advantages over existing methods in 3D samples.
Although there are some shortcomings, researchers believe that Omnipose can be a solution to various problems in bacterial cell biology.
To verify whether Omnipose can also become a multifunctional tool in other microscope-dependent biological science or even non-life science fields, the team tried to use Omnipose to analyze microscope images of Caenorhabditis elegans (C.elegans). C.elegans is an important model organism in the study of genetics , neuroscience , development and microbial behavior. C.elegans feeds on bacteria . Like some bacteria, C.elegans is also linear and can be freely twisted into various forms.
test shows that Omnipose can recognize C.elegans whether it is stretching, shrinking or other various movements. The researchers say this ability may be very useful in delayed tracking of neural studies during nematode movement.
At present, the team has disclosed Omnipose's source code, training data and models, and also provides complete documentation on installing and using Omnipose.
They expect that the excellent performance of Omnipose in different cell morphology and modes may unlock previously unavailable information from microscope images. Omnipose may even change the game rules of biological image analysis.
Paper link:
https://www.nature.com/articles/s41592-022-01639-4
Omnipose Complete document:
https://omnipose.readthedocs.io/
