Nuclei segmentation with point annotations from pathology images via self-supervised learning and co-training.

Area of Science:

• Digital pathology
• Computational imaging
• Machine learning for medical image analysis

Background:

• Accurate nuclei segmentation is vital for digital pathology, but fully-supervised methods require extensive, costly pixel-level annotations.
• Pathologists find it time-consuming and expensive to generate precise ground truth data.
• Point annotations offer a more accessible alternative for training segmentation models.

Purpose of the Study:

• To develop a weakly-supervised learning method for nuclei segmentation that utilizes only point annotations.
• To reduce the reliance on time-consuming and expensive pixel-level annotations in digital pathology.
• To improve the efficiency and accessibility of nuclei segmentation in whole slide image analysis.

Main Methods:

• Generating coarse pixel-level labels from point annotations using Voronoi diagrams and k-means clustering.
• Employing a co-training strategy with exponential moving average for refining coarse label supervision.
• Utilizing self-supervised visual representation learning by transforming hematoxylin component images to H&E stained images.

Main Results:

• The proposed method demonstrates superior performance compared to existing state-of-the-art methods on two public datasets.
• Achieved competitive segmentation accuracy relative to fully-supervised approaches.
• Visual and quantitative evaluations confirm the effectiveness of the weakly-supervised nuclei segmentation.

Conclusions:

• The developed weakly-supervised learning method effectively segments nuclei using only point annotations.
• This approach significantly lowers the annotation burden for digital pathology tasks.
• The method offers a promising, cost-effective alternative to fully-supervised nuclei segmentation.