Learning to count biological structures with raters' uncertainty.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning

Background:

• Deep learning excels at counting biological structures in microscopy images with well-labeled data.
• Weak multi-rater annotations, where experts disagree, present a significant challenge in biological structure counting.
• Aggregating multiple expert labels is ideal but often infeasible due to cost and scale.

Purpose of the Study:

• To develop a robust counting strategy for biological structures using small quantities of weakly labeled, multi-rater data.
• To refine deep learning model predictions by leveraging inter-rater disagreement.
• To improve the accuracy and confidence calibration of biological structure counts in challenging annotation scenarios.

Main Methods:

• A two-stage counting strategy was proposed: first detecting and counting structures, then refining predictions.
• The refinement step enhances the correlation between sample scores and rater agreement.
• The methodology was evaluated on a novel dataset of mouse brain fluorescence microscopy images with perineuronal nets.

Main Results:

• The proposed two-stage strategy significantly enhances biological structure counting performance.
• Improved confidence calibration was achieved by utilizing redundant information from multi-rater data.
• The method effectively addresses the challenge of limited and conflicting expert annotations.

Conclusions:

• This work presents an effective approach for counting biological structures with weak multi-rater annotations.
• Leveraging redundant information in small multi-rater datasets is crucial for improving model reliability.
• The developed methodology offers a practical solution for scenarios with budget and scale constraints in biological image analysis.