Efficient end-to-end learning for cell segmentation with machine generated weak annotations.

Area of Science:

• Computational Biology
• Biotechnology
• Image Analysis

Background:

• Deep learning excels at cell segmentation in microscopy but demands large, fully annotated datasets.
• Generating complete annotations is time-consuming and expensive, hindering widespread adoption.
• Weakly-supervised and self-supervised learning offer alternatives but often compromise accuracy.

Purpose of the Study:

• To develop a deep learning model for automated cell segmentation using a novel type of weak annotation.
• To enable accurate cell segmentation with reduced annotation effort and cost.
• To provide a practical alternative to full-supervision methods in single-cell analysis.

Main Methods:

• Designed a novel deep learning model architecture for end-to-end training.
• Utilized programmable, incomplete annotations derived from experimental data.
• Benchmarked the model on diverse public datasets (fluorescence and bright-field) and a custom dataset with machine-generated annotations.

Main Results:

• The proposed model achieved segmentation accuracy competitive with, and sometimes exceeding, state-of-the-art fully supervised methods.
• Demonstrated effectiveness across different imaging modalities and datasets.
• Validated the utility of programmable weak annotations for training accurate segmentation models.

Conclusions:

• The developed method offers a practical and efficient approach to cell segmentation using weak supervision.
• Reduces the annotation burden without sacrificing segmentation performance.
• Facilitates broader application of deep learning in single-cell analysis.