Speaker
Description
Microscopic imaging enables us to investigate cells and how they change, but since subtle changes are hard to see by eye, we need tools such as deep learning to help us see.
Here, we are combining label-free microscopy with deep learning to predict stem cell differentiation outcomes. This is highly relevant, as the differentiation process is labor intensive, costly and subject to high variability.
We have differentiated induced pluripotent stem cells towards beta-cells for the first four days of a standard protocol. On day four, we measured Cxcr4 expression as a marker for successful entrance of the Definitive Endoderm stage. We acquired phase-contrast images every hour and combined these images with Cxcr4 expression levels to re-train different pre-trained deep neural networks. The retrained models were then used to classify unseen images according to their Cxcr4 expression.
With our retrained ResNet18, we can classify experiments into high and low amount of Definitive Endoderm cells already before day four with an accuracy larger than 0.9. This enables selection of the most successful differentiation early on, saving time and money in the lab.
| Authors | Franziska Schöb*, Dag Kristian Dysthe, Valeria Sordi |
|---|---|
| Keywords | deep learning, classification, label-free microscopy, stem cell differentiation |
