Description du projet
Un meilleur suivi des gouttelettes à l’échelle microscopique
La fabrication dans le secteur des sciences de la vie exige souvent la manipulation de quantités infimes de liquides. Toutefois, les méthodes actuelles utilisées pour analyser ces microfluides peuvent s’avérer chronophages et augmenter considérablement les coûts de production. Le projet DROPTRACK, financé par l’UE, vise à utiliser la vision par ordinateur pour créer un moyen fiable et peu coûteux de suivre et d’analyser ces microfluides. Grâce à deux algorithmes d’apprentissage profond, le logiciel DROPTRACK sera en mesure de détecter et de suivre des gouttelettes individuelles à partir de vidéos expérimentales et de mesurer des quantités physiques telles que le nombre de gouttelettes, leur débit et leur fraction de remplissage. Les résultats préliminaires montrent que le logiciel peut analyser les images beaucoup plus rapidement qu’un appareil photo numérique classique, ce qui pourrait permettre des gains de temps considérables aux utilisateurs dans les domaines biologiques.
Objectif
Microfluidics technology targets droplet manipulation by confining fluids to manufacture materials for many industrial applications and life sciences. Data analysis obtained by microfluidic experiments remains a bottleneck to this day due to the lack of a reliable interface converting raw observations to informative data fast enough and at low operating costs. Many computer vision tools exist, which permit to infer useful physical information from video data but they typically target specific physical systems and lack broad applicability. Moreover, the existing tools need intensive development and fine-tuning before they can be useful in a specific scenario. Thus, such a lack of computer vision tools with broad applicability hinders the penetration of this technology for extensive microfluidics applications.
This PoC aims to develop a stand-alone, easy-to-use software (DropTrack) that can detect and track individual droplets from experimental videos. Two cutting-edge deep-learning algorithms, YOLO for droplet detection and DeepSORT for droplet tracking, will be at the software's core. DropTrack will provide trajectories of the individual droplets and measure physical quantities of interest, such as the droplet numbers, flow rates, packing fraction, etc., by analyzing videos from experiments. Our preliminary work indicates that DropTrack is capable of analyzing images much faster than the image capture rate of a typical digital camera, thereby opening entirely new prospects for real-time feedback control in experiments.
The realization of DropTrack will lead to dramatic time and resources savings in video data analysis of deformable moving objects in microfluidics. Beyond this PoC, the successful demonstration of this technology in microfluidics is expected to attract significant attention as a handy data analysis tool in other relevant fields such as multi-organism biological systems from cell aggregates to animal congregation.
Champ scientifique
- natural sciencescomputer and information sciencesdata science
- natural sciencesphysical sciencesclassical mechanicsfluid mechanicsmicrofluidics
- engineering and technologyother engineering and technologiesfood technology
- engineering and technologymaterials engineering
- natural sciencescomputer and information sciencesartificial intelligencemachine learningdeep learning
Programme(s)
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Régime de financement
HORIZON-AG-LS - HORIZON Lump Sum GrantInstitution d’accueil
16163 Genova
Italie