Many news have been published since 2 or 3 years about Artificial Intelligence (AI) and Machine Learning / Deep Learning use in picture analysis. Lastly, Microsoft published an astonishing article where they explained that this technology is so powerful that the government needs to regulate it, especially for face recognition. In HCS Pharma we are deeply involved in picture analysis and we have worked on these technologies with the help of our partners (TISBio Lille, Molecular Devices, agence dad) since 2017.
Why is it important for us?
As we are working in 3D, we need to master all the processes in 3D: 3D biological models, 3D culture with BIOMIMESYS® products, 3D pictures acquisition with ImageXpress micro confocal systems from Molecular Devices and finally, 3D pictures analysis with MetaXpress software (also from Molecular Devices).
We get nice results with MetaXpress 3D modules…
For example, you can see below a big spheroid with well contrasted cells. In this case, it’s really easy to segment cells inside, up to 50 to 60 um.
Then, we can extract relevant parameters like cell positions, intensities of fluorescence, morphological elements, co-localization, etc.
For example, the spheroid is rebuilded on the picture (on the right) with x,y positions of each cell, z distance from the plate bottom as color and size of cells representing DAPI intensity.
… but we search for a better segmentation in complex pictures
When you mix many shapes in a 3D pictures, like cells, neurites, spots or holes, it becomes more difficult to segment precisely. Indeed, because objects are occulted by other objects, the blur in Z direction is increasing (translucient medium) and, of course, precision in the Z direction are often ten times worse than in the x/y directions.
For example, look at this picture of Luhmes cells in BIOMIMESYS® Brain (more explanation here) with DAPI and FITC channels. We used WEKA to separate “area of objects”: cell nuclei in DAPI and neurites in FITC. With a short training where we label areas, WEKA classified each pixel and produced a segmentation map. Results for the whole stack (from 0 to 70 microns) is visible below:
On the left picture, it is possible to isolate 4 areas:
- Blue : mostly cell nuclei structures
- Green : mostly neurites structures
- Clear Blue : both structures
- Black : nothing
Results are really promising because WEKA also gave us “probability maps” for each structure. Theses maps could be used to filter raw pictures and help to create efficient 3D reconstructions.
To masterise machine learning in picture analysis is not easy, even if some softwares seems like “push the button and look at the results” 🙂 In fact, you must investigate many variables, many methods and always get back to biological facts. But indeed, improving segmentations is a good investment!
We are now working on teaching the software how o efficiently recognize cellular and subcellular structures in complex 3D cultures.