BoNuS: Boundary Mining for Nuclei Segmentation With Partial Point Labels.

Area of Science:

• Digital Pathology
• Computer Vision
• Biomedical Image Analysis

Background:

• Accurate nuclei segmentation is crucial for quantitative analysis in digital pathology.
• Manual annotation of nuclei is labor-intensive, time-consuming, and requires expertise.
• Automated methods are needed to overcome the limitations of manual annotation.

Purpose of the Study:

• To develop a weakly-supervised nuclei segmentation method requiring only partial point labels.
• To introduce a novel boundary mining framework (BoNuS) for simultaneous learning of nuclei interior and boundary information.
• To address the challenge of partial point labels by incorporating a nuclei detection module with curriculum learning.

Main Methods:

• A novel boundary mining loss function guides the model using pairwise pixel affinity in a multiple-instance learning framework.
• A nuclei detection module with curriculum learning is proposed to handle missing nuclei information.
• The BoNuS framework simultaneously learns nuclei interior and boundary features from limited point annotations.

Main Results:

• The proposed BoNuS method demonstrates superior performance compared to existing weakly-supervised nuclei segmentation techniques.
• Validation on MoNuSeg, CPM, and CoNIC datasets confirms the method's effectiveness.
• The approach successfully segments nuclei using significantly less annotation effort.

Conclusions:

• Weakly-supervised nuclei segmentation using partial point labels is feasible and effective.
• The BoNuS framework offers a promising solution for efficient and accurate nuclei segmentation in digital pathology.
• This method reduces the reliance on extensive manual annotation, accelerating pathological image analysis.