Generative Adversarial Networks for Morphological-Temporal Classification of Stem Cell Images

Area of Science:

• Computational biology
• Bioimage analysis
• Machine learning in life sciences

Background:

• Neural network training on biological images is often data-limited due to experimental constraints.
• Insufficient data hinders efficient learning and classification accuracy in deep learning models.
• High-throughput biological imaging requires robust methods to overcome data scarcity.

Purpose of the Study:

• To augment datasets of induced pluripotent stem cell microscopy images using generative adversarial networks (GAN).
• To enhance the classification accuracy of deep convolutional neural networks (CNNs) trained on these augmented datasets.
• To address challenges in modeling complex biological data with GANs and training on generated data.

Main Methods:

• Generated image patches of cell colonies from gray-scale microscopy images using GANs.
• Integrated generated images into real datasets to address class imbalances during neural network training.
• Employed a temporally constrained, hierarchical classification scheme incorporating domain knowledge for model learning.

Main Results:

• Achieved a 2% increase in true positive rate and F1-score compared to a standard imbalanced classification network.
• Observed greater classwise improvements in classification performance.
• Demonstrated the effectiveness of GAN-based data augmentation for biological image datasets.

Conclusions:

• Synergistic model design integrating domain knowledge is crucial for effective biological image analysis.
• GAN-based data augmentation can significantly improve neural network learning in high-throughput biological imaging.
• This approach offers a viable solution to data limitations in training deep learning models for cellular image classification.