DEEP LEARNING/ARTIFICIAL INTELLIGENCE HIGHLY PREDICTS EARLY CHANGES IN PLURIPOTENT STEM CELLS

ALEJANDRO DAMIÁN LA GRECA; ARIEL WAISMAN; ALAN MIQUEAS MÖBBS; MARIA AGUSTINA SCARAFIA; NATALIA LUCÍA SANTÍN VELAZQUE; GABRIEL NEIMAN; LUCIA NATALIA MORO; CARLOS DANIEL LUZZANI; GUSTAVO EMILIO SEVLEVER; ALEJANDRA GUBERMAN; SANTIAGO GABRIEL MIRIUKA

Mar del Plata

Congreso; LXIII Annual Meeting of the Argentine Society for Clinical Investigation (SAIC); 2018

SAIC - Sociedad Argentina de Investigacion Clinica

Thereis an ongoing revolution with the use of powerful algorithmscollectively known as machine learning. One of its branches, deeplearning (DL), allows to predict unique features based on therelation of numbers through the application of vast neural networks.Digital images, as a plain collection of numbers, are susceptible tobe processed by DL. Hence, DL can recognize image forms and classifythem. We then applied DL to phase contrast images of pluripotent stemcells (PSC) under two different models previously developed in ourlabs. In a first model of cell differentiation, naïve mouse PSC weredifferentiated to Epiblast-like cells by removing stemness factors.In a second model of cell death, human PSC were incubated withcamptothencin, a topoisomerase inhibitor, which rapid and massivelyinduces apoptosis. Both models were validated through standardassays, including real time PCR and fluorescent staining. We tookhundreds of microscopic phase contrast images in an EVOS microscopewith a 10x objective. Image processing was done in AWS using Keras asfrontend and TensorFlow as backend. We trained deep residual networksunder different settings. We found that DL training can correctlyclassify images in both models with an accuracy close to 1.Independent test on non-trained images and validation in newreplicates confirmed the high accuracy of both DL in both cellmodels. Importantly, such high accuracy was reached afterapproximately 30 minutes of cell differentiation and 3 hours aftercell death induction. In summary, we found that applying deeplearning algorithm to plain, non-stained microscopic images canreadily detect morphological changes in almost all cases andcorrectly classify them. Our findings predicts potent applications ofdeep learning/artificial intelligence in cell image detection andclassification.