Self-supervised pseudo-colorizing of masked cells.

Area of Science:

• Biomedical imaging
• Deep learning
• Computational biology

Background:

• Self-supervised learning is increasingly vital in deep learning for biomedical image analysis.
• Current methods often require extensive labeled data, limiting their application.
• Novel self-supervision objectives are needed to improve representation learning for cell analysis.

Purpose of the Study:

• To introduce a novel self-supervision objective for analyzing cells in biomedical microscopy images.
• To enhance deep learning models for cell detection through pseudo-colorization and masked image modeling.
• To evaluate the proposed method against established self-supervised learning frameworks.

Main Methods:

• Developed a novel self-supervision objective involving pseudo-colorizing masked cells using a physics-informed pseudo-spectral colormap.
• Incorporated masked image modeling by masking cell parts and training models to reconstruct them.
• Utilized hybrid deep learning models combining convolutional and vision transformer modules for cell detection.
• Compared the proposed pre-training method with contrastive learning (SimCLR), masked autoencoders (MAEs), and edge-based self-supervision.

Main Results:

• Pseudo-colorization as an approximation for semantic segmentation proved beneficial for subsequent cell detection fine-tuning.
• Masked cell part reconstruction further enriched learned representations.
• The proposed pre-training method outperformed SimCLR, MAE-like masked image modeling, and edge-based self-supervision on six diverse fluorescence microscopy datasets.
• Hybrid models demonstrated strong performance in cell detection tasks.

Conclusions:

• The proposed pseudo-colorization and masked image modeling self-supervision strategy offers a powerful approach for biomedical cell image analysis.
• This method effectively learns robust representations, improving downstream cell detection tasks.
• The approach demonstrates superior performance compared to existing self-supervised learning techniques, highlighting its potential for advancing biomedical deep learning applications.