CellTranspose: Few-shot Domain Adaptation for Cellular Instance Segmentation.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning

Background:

• Automated cellular instance segmentation accelerates biological research.
• Current generalized models fail on novel, differently distributed data.
• Retraining models requires extensive data and computational power.

Purpose of the Study:

• Develop an approach for cellular instance segmentation requiring minimal annotated data and training time.
• Address the challenge of model adaptation to novel datasets.
• Improve the efficiency and accessibility of automated segmentation tools.

Main Methods:

• Designed specialized contrastive losses for efficient leverage of few annotated samples.
• Focused on model adaptation rather than full retraining.
• Evaluated performance on novel datasets with limited annotations.

Main Results:

• 3-5 annotations yielded models that significantly mitigate covariate shift.
• Achieved accuracy matching or surpassing other adaptation methods.
• Approached the performance of fully retrained models with minimal adaptation time (minutes).

Conclusions:

• The proposed method offers a balance between model performance, computational requirements, and annotation needs.
• Enables efficient adaptation of segmentation models to new biological data.
• Facilitates broader application of automated cellular instance segmentation in research.