2D Multimodal Image Collection for Fluorescence Prediction from Transmitted Light Microscopy.

Area of Science:

• Bioimage analysis
• Machine learning in biology
• Cellular imaging

Background:

• Accurate fluorescence prediction from label-free microscopy is crucial for advancing biological research.
• Existing datasets often lack the diversity and scale needed for robust machine learning model training.

Purpose of the Study:

• To introduce the Light My Cells Database, a comprehensive open-access resource for machine learning in fluorescence prediction.
• To provide a diverse collection of microscopy images to support the development of generalizable deep learning models.

Main Methods:

• Compiled 2,574 acquisition sets and 56,984 2D microscopy images from 30 studies across 8 imaging centers.
• Paired transmitted light images (bright-field, phase contrast, DIC) with fluorescence images of subcellular structures (nucleus, mitochondria, tubulin, actin).
• Standardized images in OME-TIFF format with REMBI-compliant metadata and applied a preprocessing pipeline for harmonization.

Main Results:

• The database contains a wide diversity of biological samples, imaging modalities, and acquisition systems, reflecting real-world microscopy variability.
• Standardized images and rich metadata facilitate downstream applications like in silico labeling, segmentation, and cell profiling.
• The dataset is suitable for training and benchmarking generalizable deep learning models for label-free imaging analysis.

Conclusions:

• The Light My Cells Database is a valuable resource for advancing machine learning applications in microscopy and bioimage analysis.
• This open-access dataset will accelerate the development of AI-powered tools for label-free cellular imaging and analysis.
• Facilitates in silico labeling and cell profiling, enabling new avenues in biological discovery.