Weakly Supervised Cell Segmentation by Point Annotation.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Accurate cell segmentation is crucial for biological research.
• Fully supervised methods require extensive pixel-level annotations, which are time-consuming and labor-intensive.
• Weakly supervised learning offers a promising alternative to reduce annotation burden.

Purpose of the Study:

• To develop and evaluate weakly supervised training schemes for end-to-end cell segmentation networks.
• To achieve high-quality cell segmentation using minimal annotation, specifically a single point per cell.
• To compare the performance of proposed methods against fully supervised approaches.

Main Methods:

• Investigated three training schemes: self-training, co-training, and a hybrid approach.
• Utilized a single point annotation per cell as the training label.
• Proposed a divergence loss to prevent overfitting and a consistency loss for network consensus.
• Incorporated a human-in-the-loop strategy for enhanced annotation efficiency.

Main Results:

• Achieved high-quality cell segmentation masks comparable to fully supervised methods.
• Demonstrated significant reduction in human annotation effort.
• Validated the effectiveness of self-training, co-training, and hybrid schemes on benchmark datasets.

Conclusions:

• Weakly supervised cell segmentation using point annotations is feasible and effective.
• The proposed training schemes offer a practical solution for reducing annotation costs in cell segmentation.
• This approach enables high segmentation accuracy with substantially less manual annotation.