Double Uncertainty-Aware Learning Network for Multi-Modal Cell Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate cell image segmentation is crucial for clinical applications.
• Existing methods struggle with oversegmentation/under-segmentation due to modal variations and limited labeled data.
• Previous approaches lack global and local uncertainty perception.

Purpose of the Study:

• To develop a novel framework for accurate multi-modal cell segmentation.
• To address data and modal uncertainty in cell image analysis.
• To overcome limitations of limited labeled data in multi-modal cell segmentation.

Main Methods:

• A multi-branch image fusion module with dilation convolution, regular convolution, and channel attention.
• A transformer-based encoding strategy with token selection/enhancement based on confidence scores from a teaching network.
• A pseudo-label selection strategy to enhance unlabeled data annotation quality.

Main Results:

• The proposed framework achieved superior performance compared to fifteen existing methods on three public datasets.
• Demonstrated effective learning of valuable information for multi-modal cell segmentation.
• Successfully addressed oversegmentation and under-segmentation issues caused by uncertainty variations.

Conclusions:

• The novel framework significantly improves multi-modal cell segmentation accuracy.
• The method enhances clinical applications by improving diagnostic accuracy and decision-making reliability.
• This work offers a robust solution for cell image segmentation challenges with limited data and complex uncertainties.