A Generalized Deep Learning Pipeline for Stain-Invariant Ultrastructural Segmentation in Peripheral Nerves

Area of Science:

• Neuroscience
• Computational Biology
• Microscopy

Background:

• Automated analysis of peripheral nerve ultrastructure via electron microscopy (EM) faces challenges due to data heterogeneity.
• Variations in staining protocols and resolutions across EM datasets create domain shifts, hindering deep learning model performance.

Purpose of the Study:

• To develop a generalized segmentation pipeline for standardizing heterogeneous EM datasets.
• To mitigate domain gaps in EM data using classical image standardization techniques integrated with deep learning.

Main Methods:

• A custom pre-processing workflow involving Contrast Limited Adaptive Histogram Equalization (CLAHE) and noise suppression was implemented in ZEISS Arivis Pro.
• Standardization was applied across three distinct EM preparation domains: osmium-based Palade, lanthanide-based Uranyl-free, and low-resolution Ellisman.
• A U-Net model was trained on a composite dataset and evaluated using Intersection over Union (IoU) and Dice Similarity Coefficient (DSC).

Main Results:

• The U-Net model achieved peak internal IoU scores >0.95 for myelin and Schwann cells.
• During zero-shot inference on an independent cohort, the model maintained robust DSC scores: 0.854 for myelin and 0.597 for mitochondria.
• The pipeline successfully enabled 3D multi-organelle reconstructions from challenging, diverse EM data.

Conclusions:

• Integrating classical image standardization with deep learning effectively overcomes domain gaps in EM datasets.
• The developed pipeline supports the adoption of sustainable, non-toxic EM protocols.
• This approach provides a robust method for automated organellomics and analysis of historical clinical EM archives.