Deep Learning Method for Classifying Spatial Patterning of Early Differentiated Human Induced Pluripotent Stem Cells...

Micropattern cultures have been used to analyze early differentiation process of human induced pluripotent stem cells (hiPSCs), because the differentiated cells form segregated, sorted and radially ordered three germ layers. However, such spatial patterning of a hiPSC line with genetic abnormalities may be rather highly variable. To characterize such abnormal patterns, we present a CNN structure and global average pooling for classifying micropattern images of early differentiated hiPSCs. In this structure, the downsampling is performed successively to reduce feature maps as half and double the filters at every stage. The global average pooling reduces the impact of over-fitting during training to obtain optimal classification. Seven classes of ectoderm al cell images were used to represent the status or type of early hiPSC differentiation with genetic abnormalities such as trisomy, uniparental disomy and loss of heterozygosity. The accuracy achieved 81.4%, indicating the proposed method is useful to analyze the differentiation of hiPSCs with genetic abnormalities.