[A region-level contrastive learning-based deep model for glomerular ultrastructure segmentation on electron...

Area of Science:

• Electron microscopy imaging analysis
• Computational pathology
• Machine learning for biomedical imaging

Context:

• Accurate segmentation of glomerular ultrastructures is crucial for understanding kidney disease.
• Existing methods struggle with the scarcity of labeled electron microscopy data.
• Self-supervised learning offers a promising avenue to leverage large unlabeled datasets.

Purpose:

• To introduce Ultrastructural Region Contrast (USRegCon), a novel region-level self-supervised contrastive learning method.
• To improve glomerular ultrastructure segmentation performance on electron microscope images.
• To overcome the limitations of labeled data scarcity in biomedical image analysis.

Summary:

• USRegCon pre-trains models using unlabeled electron microscopy images in three steps: adaptive region division based on semantic similarity, extraction of grayscale and deep semantic region representations, and joint optimization using grayscale and semantic loss functions.
• The method effectively minimizes intra-region grayscale differences and maximizes inter-region differences, while enhancing the similarity of semantically similar regions in the representation space.
• This approach enables the model to learn robust region representations from large unlabeled datasets.

Impact:

• USRegCon achieved high Dice coefficients for segmenting basement membrane (85.69%), endothelial cells (74.59%), and podocytes (78.57%) on the GlomEM dataset.
• The method demonstrates superior performance compared to existing self-supervised learning techniques and approaches fully-supervised methods.
• USRegCon effectively addresses the challenge of limited labeled data, enhancing deep model performance for ultrastructure recognition and segmentation.