Weakly Supervised Nuclei Segmentation with Point-Guided Attention and Self-Supervised Pseudo-Labeling.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Manual nuclei segmentation is labor-intensive.
• Weakly supervised methods face challenges with adjacent nuclei and noisy pseudo-labels.

Purpose of the Study:

• To develop a precise nuclei segmentation method using point supervision.
• To address challenges in segmenting adjacent nuclei and mitigating pseudo-label noise.

Main Methods:

• A Gaussian kernel mechanism for multi-scale center-point prediction to separate adjacent nuclei.
• A point-guided attention mechanism to reduce noise from pseudo-labels.
• An exponential moving average (EMA) and k-means clustering-based label updating mechanism to enhance pseudo-label quality.

Main Results:

• Achieved state-of-the-art performance on three public datasets across multiple metrics.
• Demonstrated effective separation of adjacent cell nuclei.
• Significantly reduced the impact of noise from inaccurate pseudo-labels.

Conclusions:

• The proposed method offers precise nuclei segmentation with reduced annotation costs.
• Facilitates large-scale dataset training and promotes automated analysis in clinical settings.
• Reduces reliance on clinical experts for segmentation tasks.