One-shot cell segmentation via learning memory query: Towards universal solution without active tuning.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Cell Biology

Background:

• Cell segmentation is crucial for disease analysis and drug development.
• Current methods are often image-specific and require extensive manual adjustments, increasing labor.
• There is a need for adaptable and efficient cell segmentation tools.

Purpose of the Study:

• To introduce Mimic, a novel framework for automated cell segmentation.
• To develop a generalist cell segmentation model that requires minimal user intervention.
• To improve the speed and accuracy of quantitative cell analysis.

Main Methods:

• Mimic utilizes a Query-and-Answer (Q&A) mechanism for single-step cell segmentation.
• The model learns from a few example prompts, enabling adaptation to new cell types without retraining.
• The framework was evaluated on 12 diverse public datasets.

Main Results:

• Mimic achieved state-of-the-art performance across various imaging techniques, cell types, and staining methods.
• The model outperformed existing generalist cell segmentation tools like Cellpose and Stardist.
• Mimic demonstrated superior adaptability, requiring no additional training for new cell types.

Conclusions:

• Mimic offers an efficient and accurate solution for cell segmentation in biomedical research.
• The Q&A approach and prompt-based learning reduce labor and enhance model generalizability.
• This framework has the potential to accelerate quantitative analysis in biological and medical studies.