Scellseg: A style-aware deep learning tool for adaptive cell instance segmentation by contrastive fine-tuning.

Area of Science:

• Computational Biology
• Bioimage Analysis
• Machine Learning for Microscopy

Background:

• Deep learning cell segmentation is vital for analyzing large biological datasets.
• Limited annotated data hinders specialist algorithm development, while generalist models lack experiment-specific adaptation.

Purpose of the Study:

• To introduce Scellseg, an adaptive pipeline for efficient and accurate cell segmentation.
• To address the challenge of limited annotated data in deep learning-based cell segmentation.

Main Methods:

• Utilizes a style-aware pre-trained model combined with contrastive fine-tuning on unlabeled data.
• Employs an adaptive pipeline for self-improving segmentation models.
• Develops a graphical user interface for user-friendly specialization.

Main Results:

• Achieves state-of-the-art transferability in average precision and Aggregated Jaccard Index across diverse datasets.
• Demonstrates that model performance plateaus after fine-tuning on approximately eight images.
• Shows that Scellseg can create specialist models with minimal manual annotation effort.

Conclusions:

• Scellseg offers a powerful and efficient solution for cell segmentation in biological research.
• The adaptive pipeline and few-shot learning capability significantly reduce the need for extensive manual annotation.
• The user-friendly interface democratizes the creation of customized cell segmentation models for biologists.