Omni-Seg: A Scale-Aware Dynamic Network for Renal Pathological Image Segmentation.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Renal pathological image segmentation is complex due to significant variations in object scales.
• Existing multi-network approaches are resource-intensive and do not capture spatial relationships between tissue types.

Purpose of the Study:

• To introduce Omni-Seg, a single, scale-aware dynamic neural network for multi-object and multi-scale renal pathological image segmentation.
• To address the challenge of segmenting heterogeneous tissue types at different resolutions within a unified framework.

Main Methods:

• Development of a novel scale-aware controller for generalizing dynamic neural networks to multi-scale segmentation.
• Implementation of semi-supervised consistency regularization with pseudo-labels to model inter-scale correlations.
• Training a single network on 150,000 human kidney pathological image patches across six tissue types and three resolutions.

Main Results:

• Omni-Seg successfully performs multi-object and multi-scale segmentation using a single neural network.
• The model demonstrates superior scale-aware generalization, performing effectively on mouse kidney images without retraining.
• Achieved enhanced segmentation performance validated by human visual assessment and image-omics analysis.

Conclusions:

• Omni-Seg offers an efficient and effective solution for renal pathological image segmentation, overcoming scale heterogeneity challenges.
• The proposed scale-aware controller and semi-supervised learning paradigm advance the field of medical image analysis.
• This single-network approach provides a robust foundation for analyzing kidney pathology and related image-omics data.