Cell-DINO: Self-supervised image-based embeddings for cell fluorescent microscopy.

Area of Science:

• Computational Biology
• Bioimage Analysis
• Machine Learning

Background:

• Accurate quantification of cellular morphology is crucial for single-cell biological studies.
• Existing computer vision methods for cell morphology analysis often require extensive manual annotation.
• Measuring cell morphology at scale remains a significant challenge in biological research.

Purpose of the Study:

• To evaluate the efficacy of DINOv2, a self-supervised vision transformer, in learning cellular morphology representations without supervision.
• To develop and assess Cell-DINO models for cell phenotyping tasks.
• To demonstrate the utility of Cell-DINO in improving performance, particularly in low-annotation scenarios.

Main Methods:

• Application of DINOv2, a self-supervised learning algorithm, to extract features from cellular images.
• Development of Cell-DINO models by applying DINOv2 to cell phenotyping challenges.
• Comparative analysis of Cell-DINO models against supervised and other self-supervised baselines on diverse imaging datasets.

Main Results:

• DINOv2 effectively learns rich cellular morphology representations without manual annotations.
• Cell-DINO models show superior performance compared to supervised and other self-supervised methods across various tasks.
• Significant performance gains were observed in low-annotation regimes, e.g., 70% improvement in protein localization classification with 1% annotations.

Conclusions:

• Cell-DINO models offer a powerful, annotation-efficient approach for cell phenotyping.
• This method can facilitate the study of biological variation, including single-cell heterogeneity and experimental condition relationships.
• Cell-DINO represents a valuable tool for advancing image-based biological discovery.