DeephESC 2.0: Deep Generative Multi Adversarial Networks for improving the classification of hESC

Area of Science:

• Biotechnology
• Machine Learning
• Stem Cell Biology

Background:

• Human embryonic stem cells (hESC) are crucial for disease modeling and regenerative medicine.
• Current manual annotation of hESC videos is labor-intensive and time-consuming.
• Automated analysis is needed for efficient quantification of hESC states in research.

Purpose of the Study:

• To introduce DeephESC 2.0, an automated system for classifying hESC in videos.
• To develop a non-invasive method for hESC image analysis.
• To improve the efficiency and accuracy of hESC state identification.

Main Methods:

• Utilized Generative Multi Adversarial Networks (GMAN) to create synthetic hESC images.
• Implemented a hierarchical classification system with Convolution Neural Networks (CNN) and Triplet CNNs.
• Classified hESC images into six distinct categories without chemical staining.

Main Results:

• Achieved 93.23% accuracy in classifying hESC images, surpassing state-of-the-art methods by over 20%.
• Demonstrated that training with synthetic images improved classifier performance to 94.46%.
• Generated high-quality synthetic images evaluated using SSIM, PSNR, and p-value metrics.

Conclusions:

• DeephESC 2.0 offers a highly accurate and efficient automated solution for hESC analysis.
• The system significantly reduces manual labor, saving hundreds of hours.
• Synthetic data generation enhances model performance and dataset augmentation capabilities.