Iterative pseudo balancing for stem cell microscopy image classification.

Area of Science:

• Computational Biology
• Machine Learning
• Bioinformatics

Background:

• Deep neural networks (DNNs) struggle with limited, imbalanced biological datasets, leading to overfitting and reduced accuracy.
• Manual annotation of biological datasets is time-consuming and expensive, hindering research.
• Semi-supervised models are needed to reduce reliance on large, manually annotated datasets.

Purpose of the Study:

• To develop a semi-supervised deep learning model for classifying stem cell microscopy images.
• To address challenges of limited and imbalanced biological datasets.
• To improve the accuracy and efficiency of neural network training in biological imaging.

Main Methods:

• Introduction of Iterative Pseudo Balancing (IPB) for on-the-fly dataset balancing.
• Utilizing a student-teacher meta-pseudo-label framework for semi-supervised learning.
• Incorporating multi-scale patches from multi-label images to capture local and global features.

Main Results:

• The proposed deep neural network achieved a statistically significant 3% increase in classification accuracy over the baseline.
• Iterative Pseudo Balancing (IPB) effectively balanced datasets during training.
• The integration of multi-scale image features enhanced learning effectiveness and efficiency.

Conclusions:

• Novel application of pseudo-labeling in data-limited biological settings.
• Demonstrated the importance of utilizing all available image features for semi-supervised network performance.
• The proposed methods reduce the need for manual annotation, accelerating scientific research in cellular imaging.