Self-Supervised Voxel-Level Representation Rediscovers Subcellular Structures in Volume Electron Microscopy

Area of Science:

• * Computational biology
• * Machine learning
• * Biological imaging

Background:

• * Unsupervised representation learning is crucial for analyzing large biological imaging datasets.
• * Existing methods often focus on cropped images, lacking robust models for whole cell volumes.
• * A general model mapping every voxel to a latent space for unsupervised segmentation is needed.

Purpose of the Study:

• * To develop a robust, general model for voxel-level representation learning in whole cell volumes.
• * To achieve unsupervised segmentation of complete cells using learned representations.
• * To improve upon existing methods by separating latent space into semantic and transformational components.

Main Methods:

• * Employed variational auto-encoder and metric learning for voxel-level representation.
• * Introduced a novel approach to separate latent space into semantic and transformational subspaces.
• * Utilized the semantic representation for unsupervised segmentation.

Main Results:

• * Achieved self-supervised voxel-level representation and unsupervised segmentation for complete cells.
• * Demonstrated that the learned semantic representation visually distinguishes major subcellular components.
• * Showed the semantic subspace is more transformation-invariant than other latent subspaces.
• * Unsupervised segmentation successfully rediscovered major cellular classes and dissected unspecified areas by texture.

Conclusions:

• * The proposed method provides the first self-supervised voxel-level representation and unsupervised segmentation for complete cells.
• * Separating latent space into semantic and transformational components enhances segmentation accuracy.
• * The model outperforms baseline methods significantly in unsupervised cell segmentation.