Cellpose 2.0: how to train your own model.

Area of Science:

• Computational Biology
• Image Analysis
• Machine Learning

Background:

• Pretrained neural networks offer general biological segmentation but lack user-specific adaptation.
• Suboptimal performance occurs when test images differ significantly from training data.

Purpose of the Study:

• Introduce Cellpose 2.0, a package with diverse pretrained models and a human-in-the-loop pipeline.
• Enable rapid prototyping and customization of biological segmentation models.
• Improve segmentation accuracy and adaptability for varied biological image datasets.

Main Methods:

• Utilized an ensemble of diverse pretrained models within Cellpose 2.0.
• Developed a human-in-the-loop pipeline for efficient model fine-tuning.
• Compared performance of fine-tuned models with models trained on extensive datasets.

Main Results:

• Fine-tuning Cellpose models with 500-1,000 regions of interest (ROI) approached performance of models trained on up to 200,000 ROI.
• Human-in-the-loop approach reduced annotation needs to 100-200 ROI while maintaining segmentation quality.
• Cellpose 2.0 provides tools like an annotation GUI, model zoo, and pipeline for user adoption.

Conclusions:

• Cellpose 2.0 significantly reduces the data annotation burden for custom biological image segmentation.
• The human-in-the-loop pipeline enhances model adaptability and performance for specific user needs.
• Cellpose 2.0 democratizes advanced biological image segmentation through accessible software tools.