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

Recent clinical trial data have shown that cell therapy holds curative potential for type-I diabetes. The required cell differentiation process exhibits substantial variability, even among clones of induced pluripotent stem cells (iPSC). We apply an EfficientNet-V2-S model to obtain a novel early prediction for the outcome of the differentiation process from patient-derived iPSC to β-cells using only phase-contrast images. In contrast to many established data sets in medical imaging, human experts struggle to predict the ground truth in this type of images. Therefore, to gain critical insight into the learned features, we applied layer-wise relevance propagation (LRP), and Fourier-based frequency analysis. LRP analysis shows that the negative class contains more hole-like regions between cells. We show that the model learns features related to patch variance, and that patch-normalization improves patch accuracies for times after cell washing with less cell debris.