Instance-Aware Pseudo-Labeling and Class-Focused Contrastive Learning for Weakly Supervised Domain Adaptive...

Area of Science:

• Neuroscience
• Cell Biology
• Machine Learning

Background:

• Accurate segmentation of mitochondria in electron microscopy (EM) images is crucial for biological and neuroscience research.
• Unsupervised domain adaptation (UDA) methods face performance limitations in practical applications due to domain shifts.
• Annotation costs for detailed segmentation in EM images are prohibitively high.

Purpose of the Study:

• To develop an annotation-efficient segmentation method for mitochondria in EM images.
• To investigate weakly supervised domain adaptation (WDA) using sparse point labels for improved accuracy.
• To reduce the annotation effort and expert knowledge required for EM image segmentation.

Main Methods:

• Introduced a multitask learning framework combining segmentation and center detection.
• Employed a novel cross-teaching mechanism and class-focused cross-domain contrastive learning.
• Implemented segmentation self-training with an instance-aware pseudo-label (IPL) selection strategy for reliable pseudo-labeling.

Main Results:

• The proposed WDA method significantly outperforms existing UDA and WDA techniques on challenging datasets.
• The method substantially narrows the performance gap compared to fully supervised approaches.
• Achieved notable improvements over other UDA techniques even in an unsupervised setting.

Conclusions:

• Weakly supervised domain adaptation with sparse point labels is a viable and effective strategy for mitochondria segmentation.
• The proposed multitask learning and instance-aware pseudo-labeling approach enhances segmentation accuracy and efficiency.
• This method offers a practical solution for large-scale EM image analysis in neuroscience and cell biology.