Using Deep Learning Techniques in Daily Image Processing at Light Microscopy Facility

Session

Poster Session

Authors

Martin Čapek (1), Michaela Blažíková (1), Jan Valečka (1), Ivan Novotný (1), Helena Chmelová (1), Jiří Černý (1), Ondrej Horváth (1)

Affiliations

1. Light Microscopy, Institute of Molecular Genetics of the Czech Academy of Sciences, Prague

Keywords

image segmentation, tracking, deep learning, light microscopy

Abstract text

Today, deep learning (DL) is widely used for image processing tasks such as deconvolution, denoising, image-to-image translation, object detection, segmentation, tracking. The Institute of Molecular Genetics' Light Microscopy Facility has successfully integrated these techniques, and we'll now present practical examples of DL's application in image segmentation and tracking.

We employed the popular and freely available StarDist [1] DL project for segmenting c-Fos expression spots in mouse brain slice images captured via widefield transmission microscopy (Fig. 1). StarDist was trained on subpictures from the slices and applied to the entire images. Fig. 1 shows the original data with color-coded labels using a '3-3-2 RGB' look-up table in ImageJ/Fiji, resulting in various colors. Fig. 1A focuses on a zoomed-in area from the upper left part of the slice, while Fig. 1B displays the corresponding labeled spots. Fig. 1C is a heatmap illustrating spot distribution, with warmer colors indicating higher spot density. Creating the heatmap involved a custom ImageJ/Fiji plugin [2].

Cellpose [3], another valuable DL method, was used to segment nuclei in Fig. 2A, which depicts DAPI stained nuclei in a mouse Langerhans islet captured with a Z.1 Zeiss Lightsheet microscope. In Fig. 2B, the segmented nuclei are labeled, and Fig. 2C shows the complete 3D stack with segmented nuclei. This segmentation aims to count nuclei in Langerhans islets and compare them to the expected DNA content.

Omnipose [4], a project based on Cellpose, specializes in bacterial species segmentation. We applied it to segment bacteria in time-lapse images (mCherry) captured over 300 minutes using Andor Dragonfly spinning disk fluorescence in widefield mode, as depicted in four time points (Fig. 3A,B). The upper row displays the original data with inverted intensity values, while the middle row shows the segmented data. The images were further analyzed using DeLTA 2.0 [5], deep learning-based image processing pipeline for segmenting and tracking single cells in time. A pretrained tracking model was used to explore bacterial division-related phenomena (Fig. 3C). 

MitoSegNet [6], a suite for mitochondrial morphology analysis, was customized using our own mitochondrial yeast image datasets [7]. Fig. 4A and C show mitochondrial networks for control and treated samples, while Fig. 4B and D display their respective segmentations. Fig. 4E presents tables with morphological parameters, highlighting that control data lacks clustering, whereas peroxide-treated data exhibits strong parameter clustering.

These examples highlight challenging segmentation and tracking scenarios: Fig. 1 has variable spot intensities and an uneven background. Fig. 2 shows blurred 3D data with touching objects. Fig. 3 displays closely touching species, and Fig. 4 features diverse mitochondria. Straight fragments connect, while round ones tend to separate. We also exemplify using various DL projects for various scientific projects.

Fig. 1 courtesy of Dr. Helena Janíčková, Lab Neurochemistry, Inst Physiology of the Czech Academy of Sciences, Prague

Fig. 2 courtesy of Dr. David Habart, Lab Pancreatic Islets, Inst Clinical and Experimental Medicine (IKEM), Prague

Fig. 3 courtesy of Dr. Ondřej Černý, Lab Infection Biology, Inst Microbiology of the Czech Academy of Sciences, Prague

Fig. 4 courtesy of Dr. Jana Vojtová, Lab Regulation of Gene Expression, Inst Microbiology of the Czech Academy of Sciences, Prague

References

[1] https://github.com/stardist/stardist/

[2] https://github.com/LMCF-IMG/Protein_Expression

[3] https://github.com/mouseland/cellpose

[4] https://github.com/kevinjohncutler/omnipose

[5] https://gitlab.com/dunloplab/delta

[6] https://github.com/mitosegnet

[7] https://github.com/LMCF-IMG/Morphology_Yeast_Mitochondria