DEEP LEARNING POWERED IDENTIFICATION AND CLASSIFICATION OF EARLY DIFFERENTIATED GERM LAYERS FROM PLURIPOTENT STEM CELLS

Author

Sakib Mohammad, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2025

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

First Advisor

Chowdhury, Farhan

Abstract

Pluripotent stem cells can differentiate into all three germ layers: ecto-, meso-, and endoderm. These germ layers eventually give rise to all tissues and organs in the body, making it crucial to track their differentiation in personalized regenerative medicine. Rapid identification and characterization of lineage commitment, with single cell resolution, in response to physical and chemical cues remain challenging. Therefore, it limits high-throughput screening for lineage specification to determine differentiation efficiency. In this work, we present a deep learning (DL) based approach to identify and classify early differentiated germ layer cells derived from pluripotent mouse embryonic stem cells (mESCs). We utilized two transgenic stem cell lines called OGTR1 and OGR1 and differentiated these into three germ layers. We validated the lineage identity through the upregulation of marker genes: Sox1 for ectoderm, Gata6 for endoderm, and endogenous expression of Brachyury (T) for mesoderm. After that, we utilized two image modalities, namely, phase-contrast and nucleus images, and developed convolutional neural networks (CNNs), including InceptionV3 and ResNet50, to obtain classification accuracies of up to 97% for phase-contrast images and 90% for nucleus images from single cells and cell clusters. In addition, for a more visually expressive result, we implemented Attention UNet for image segmentation and achieved a mean intersection over union (mIoU) score of 61% for the phase-contrast and 69% for nucleus images. We employed Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-Agnostic Explanations (LIME) to visualize model predictions to find out important areas of the cells necessary for class prediction, enhancing the transparency of the prediction. Finally, we evaluated the impact of dataset size, image augmentation intensity, and image bit-depth on prediction confidence. Our work systematically integrates DL with high-resolution optical imaging and molecular validation (of the gene expression) that offers a powerful tool for rapid and high-throughput identification of germ layer-specific single cells, thus advancing applications in tissue engineering andregenerative medicine.