Contrastive Learning for Mitochondria Segmentation.

Area of Science:

• Neuroscience
• Biomedical Imaging
• Computer Vision

Background:

• Accurate mitochondria segmentation in electron microscopy (EM) images is crucial for neuroscience research.
• Challenges include image degradation, diverse mitochondrial morphologies, noise, artifacts, and overlapping sub-cellular structures.

Purpose of the Study:

• To develop a novel contrastive learning framework to improve mitochondria segmentation.
• To enhance feature representation by focusing on challenging image examples.

Main Methods:

• A point sampling strategy to identify representative pixels from difficult training examples.
• A pixel-wise label-based contrastive loss incorporating similarity and consistency terms.
• The similarity term promotes intra-class pixel similarity and inter-class pixel separability in feature space.
• The consistency term improves the 3D model's sensitivity to temporal image variations.

Main Results:

• The proposed framework demonstrates effectiveness on both MitoEM and FIB-SEM datasets.
• Achieved state-of-the-art or comparable results compared to existing methods.
• Improved segmentation accuracy by learning robust feature representations from hard examples.

Conclusions:

• The novel contrastive learning framework offers an effective solution for challenging mitochondria segmentation tasks in EM.
• The method enhances segmentation performance through improved feature learning from difficult image data.
• This approach holds promise for advancing quantitative analysis in neuroscience and cell biology.