uniDINO: Assay-independent feature extraction for fluorescence microscopy images

Area of Science:

• Cellular imaging
• Bioimage analysis
• Quantitative biology

Background:

• High-content imaging (HCI) extracts quantitative features from microscopy images to characterize cellular states.
• Developing generalizable feature extraction models for HCI is difficult due to image heterogeneity (e.g., channel count, cell type, assay conditions).

Purpose of the Study:

• Introduce uniDINO, a generalist feature extraction model designed for fluorescence microscopy images with an arbitrary number of channels.
• Address the challenge of analyzing heterogeneous HCI datasets.

Main Methods:

• Trained uniDINO on over 900,000 single-channel images from diverse experimental contexts.
• Generated multi-channel image embeddings by concatenating single-channel features.
• Validated uniDINO performance across varied datasets.

Main Results:

• uniDINO outperforms traditional computer vision methods and natural image transfer learning.
• Demonstrated interpretability through channel attribution.
• Achieved superior performance on diverse microscopy datasets.

Conclusions:

• uniDINO provides an out-of-the-box, computationally efficient solution for fluorescence microscopy feature extraction.
• The model has the potential to significantly accelerate HCI dataset analysis.
• uniDINO offers a generalizable approach for handling diverse microscopy image data.