Masked Image Modeling for Generalizable Organelle Segmentation in Volume EM.

Area of Science:

• Cell Biology
• Computational Biology
• Image Analysis

Background:

• Accurate organelle segmentation in electron microscopy (EM) volumes is crucial for understanding cellular architecture.
• Deep learning methods for segmentation require extensive annotations, which are often scarce.
• Masked image modeling (MIM) pretraining leverages unlabeled data but can be inefficient for EM due to overlooked structural patterns.

Purpose of the Study:

• To develop a more efficient pretraining strategy for EM organelle segmentation.
• To improve the reliability and accuracy of deep learning models in the absence of abundant annotations.
• To introduce OrgMIM, a novel dual-branch MIM framework tailored for EM data.

Main Methods:

• Proposed OrgMIM, a dual-branch MIM framework with complementary masking strategies.
• One branch uses static structural priors and visual foundation models for affinity map generation.
• The other branch employs dynamic reconstruction feedback and self-guidance for loss map computation.
• Introduced cross-branch consistency regularization for robust representation learning.
• Constructed IsoOrg-1K, a large-scale, organelle-centric 3D EM dataset for pretraining.

Main Results:

• OrgMIM demonstrated superior performance on three public EM datasets.
• Pretraining OrgMIM on IsoOrg-1K significantly boosted segmentation accuracy (mIoU by 28.78%) compared to training from scratch.
• The framework effectively captures subcellular semantics and organelle-specific features.

Conclusions:

• OrgMIM offers an efficient and effective pretraining strategy for EM organelle segmentation.
• The proposed method enhances deep learning model performance, particularly in low-annotation scenarios.
• The IsoOrg-1K dataset and OrgMIM framework advance research in 3D EM image analysis.