S2L-CM: Scribble-supervised nuclei segmentation in histopathology images using contrastive regularization and...

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Deep learning excels in pathology nuclei segmentation but requires extensive manual labeling.
• Supervised learning demands significant effort for ground truth data generation.
• Weakly supervised learning offers a solution with sparse annotations but often yields lower performance.

Purpose of the Study:

• To propose S2L-CM, a scribble-supervised nuclei segmentation framework.
• To reduce the need for manual ground truth labels in nuclei segmentation.
• To enhance segmentation performance using refined pseudo-labels.

Main Methods:

• Leveraging self-generated pseudo-labels from sparse scribble annotations for model training.
• Utilizing multiscale contrastive regularization for pseudo-label refinement.
• Employing pixel-level multiple-instance learning to improve segmentation accuracy.

Main Results:

• S2L-CM demonstrates effectiveness and robustness across four nuclei datasets.
• The proposed method achieves competitive performance compared to state-of-the-art techniques.
• Successful nuclei segmentation achieved without requiring full ground-truth labels.

Conclusions:

• S2L-CM offers an efficient approach to nuclei segmentation by minimizing manual annotation effort.
• The framework shows promise for practical applications in digital pathology.
• Further research can build upon this method for advanced weakly supervised segmentation tasks.