Unsupervised content-preserving transformation for optical microscopy.

Area of Science:

• Biomedical imaging
• Computational microscopy
• Artificial intelligence in science

Background:

• Deep learning and open-access imaging data offer solutions for computational image transformation in biomedical research.
• Supervised deep learning methods require extensive, error-prone data annotation, limiting their broad application.
• Current limitations hinder the full potential of AI in optical microscopy.

Purpose of the Study:

• To develop an unsupervised deep learning framework for optical microscopy image transformation.
• To overcome the limitations of supervised learning by removing the need for paired training data.
• To enable wider adoption of AI in biomedical imaging research, even in data-scarce situations.

Main Methods:

• Proposed an unsupervised image transformation model named Unsupervised content-preserving Transformation for Optical Microscopy (UTOM).
• Introduced a saliency constraint to preserve image content during transformation.
• The model learns mappings between image domains without requiring paired training data.

Main Results:

• UTOM demonstrated effective performance in various biomedical image transformation tasks.
• Successful applications include in silico histological staining, fluorescence image restoration, and virtual fluorescence labeling.
• Quantitative evaluations confirmed stable, high-fidelity image transformations across diverse imaging conditions and modalities.

Conclusions:

• UTOM facilitates the use of deep learning in optical microscopy without paired data.
• The framework avoids distortions, preserving crucial image content.
• This approach is expected to drive a paradigm shift in AI model training for biomedical imaging.