Deep Learning for Predicting Stem Cell Efficiency for use in Beta Cell Differentiation

Abstract

Recent clinical trial data have shown that cell therapy holds curative potential for type-1 diabetes, however the large amounts of lab-grown cells required is a substantial bottleneck. The necessary cell differentiation process exhibits substantial variability, even among clones of induced pluripotent stem cells generated from the same patient. In contrast to many established data sets in medical imaging, human experts struggle to see the difference between highly- and lowly-efficient cell clones. We therefore propose an image-based deep learning model to guide the selection of the most efficient stem cell clones. We apply different deep learning model architectures to learn the morphological differences between good and bad stem cell clones and classify them based on only phase-contrast imaging. To gain critical insight into the learned features and enhance trust in our model, we use layer-wise relevance propagation (LRP), and Fourier-based frequency analysis. Using an EfficientNet-V2-S model, we obtain a novel early prediction for the outcome of the differentiation process from patient-derived stem cell to {beta}-cells using only phase-contrast images. Clone level accuracy is 76.7 % at 24 hours and 96.7 % at 53 hours after start of differentiation. The LRP-generated attribution maps and structure factor analysis show that the structure of the cell population is an important predictive feature. All in all, we present a highly predictive model for successful stem cell differentiation from phase-contrast images, which learns biologically relevant features. This study is a proof-of-concept that deep learning combined with label-free imaging can guide selection of induced pluripotent stem cell clones, thereby reduce cost of {beta}-cell production, and bring curative treatments one step closer to the clinic.