WeakMitoSAM: competitive prompt aggregation for point-supervised mitochondria segmentation in electron microscopy...

Area of Science:

• Cell Biology
• Biomedical Imaging
• Machine Learning

Background:

• Mitochondrial morphology is vital for understanding cell metabolism and disease.
• Accurate segmentation in electron microscopy (EM) images is crucial but challenging.
• Current deep learning methods struggle with the labor-intensive annotation process and sparse mitochondrial distributions.

Purpose of the Study:

• To develop a weakly supervised framework for high-precision mitochondria segmentation in EM images using sparse point annotations.
• To address the limitations of fully supervised and existing weakly supervised methods in EM data.

Main Methods:

• Proposed WeakMitoSAM, a novel weakly supervised framework utilizing sparse point annotations.
• Introduced competitive aggregation of multiple prompts with Bias-augmented Softmax for robust pseudo-label generation.
• Specialized Segment Anything Model (SAM) for mitochondrial ultrastructures using low-rank adaptation for efficient domain adaptation.

Main Results:

• WeakMitoSAM achieved state-of-the-art performance in point-supervised segmentation tasks across four public EM datasets.
• The method demonstrated superior performance compared to several fully supervised benchmarks.
• Achieved high-fidelity pseudo-labels from sparse point priors, effectively handling semantic ambiguities and background noise.

Conclusions:

• WeakMitoSAM offers an efficient and robust solution for large-scale mitochondrial morphofunctional analysis.
• The framework overcomes the bottleneck of dense annotations in EM image segmentation.
• Demonstrates the potential of weakly supervised learning with sparse annotations for complex biological image analysis.