As you may know, in HCS Pharma we strongly believe that the critical point of the human health evolution in next years is the in vitro 3D cell culture, especially for complex diseases like cancer . It’s why we work hard on our exclusive BIOMIMESYS® technology. As experts in HCS and cell imaging, we must master all the process : 3D biological models, 3D cell culture, volumetric pictures acquisition, 3D reconstruction and segmentation of cells and ECM compounds, 3D parameters extraction and, of course, biological interpretations. As explained in our VisuAI R&D project, the 3D reconstruction and segmentation step is not simple.
With the extraordinary automated microscopes of our partner Molecular Devices (ImageXpress® Micro Confocal High-Content Imaging System) we get very accurate 3D pictures from 96-well or 384-well plates. The reconstruction of these pictures to extract reliable parameters (like EC50 for drug discovery) is difficult because it needs many hypotheses. Don’t forget we work in a translucent medium! Of course, we are also in “HCS” conditions, means many pictures and fast acquisition, which is very different than medical pictures where automated 3D reconstruction exist .
To address these issues, we imagined the VisuAI R&D project in 2018 and we constructed a first research program with great partners (ISEN engineering school, BioMEMS team (IEMN) and TISBIO team of Université de Lille) focused on Artificial Intelligence uses in 3D complex reconstruction.
A PhD student begun to work on the subject at the end of 2019 and we get now very interesting results, soon published. Briefly, the first step was to use Machine Learning algorithms and compare them altogether and with humans expertise. One specific difficulty was also to find the right key comparators, as segmentation is not perfect, even made by humans.
We have worked with HeLa cells seeded at 25.000 cells per well in BIOMIMESYS oncology. The 3D pictures are made with our ImageXpress® Micro Confocal High-Content Imaging System with DAPI fluorochrome. We find good results in nuclear segmentation even with high level of noise in pictures, with a “reasonable” execution time for our computing platform. We also discover an effect of the Z distance to objective, on the robustness of algorithms. The “blur quantity” on picture, due to the translucent medium is related to segmentation performance.
Next step is to work on Deep Learning algorithms  to overcome the Z-stack problem and the multi-dataset problem and working on a robust architecture that take into consideration the different layers properties.
 Of Extracellular Matrix, Scaffolds, and Signaling: Tissue Architecture Regulates Development, Homeostasis, and Cancer. Annual Review of Cell and Developmental Biology 22, 287–309 (2006).
 Bücking, T. M. et al. From medical imaging data to 3D printed anatomical models. PLOS ONE 12, e0178540 (2017).
 Moen, E. et al. Deep learning for cellular image analysis. Nat Methods 16, 1233–1246 (2019).