A Deep Learning-Based Approach for Glomeruli Instance Segmentation from Multistained Renal Biopsy Pathologic Images.

Area of Science:

• Nephropathology
• Computational Pathology
• Medical Image Analysis

Background:

• Accurate glomeruli segmentation is crucial for automated renal biopsy analysis.
• Glomerular pathology varies significantly, necessitating models that handle diverse staining and lesions.
• Current methods require robust segmentation and classification for various glomerular changes.

Purpose of the Study:

• To develop and validate a robust model for glomeruli instance segmentation and classification in multi-stained renal biopsy images.
• To evaluate the model's performance across different glomerular pathologies and staining techniques.
• To establish an efficient tool for initial glomerular histologic analysis.

Main Methods:

• Utilized a dataset of pathologic images from renal biopsy slides stained with three special methods.
• Employed a Cascade Mask region-based convolutional neural network (R-CNN) architecture.
• Trained the model to segment and classify glomeruli into three categories: normal (GN), global sclerosis, and other lesions.

Main Results:

• The model achieved high F1 scores for total glomeruli segmentation (0.914 in snapshot, 0.940 in whole-slide).
• Performance for specific categories included GN (0.896/0.839), global sclerosis (0.681/0.806), and other lesions (0.756/0.753).
• Glomeruli with Normal Structure (GN) showed the best segmentation performance across both datasets.

Conclusions:

• The developed model efficiently and robustly segments and classifies multi-stained glomeruli.
• This AI tool serves as a foundational step for more in-depth glomerular histologic analysis.
• The model demonstrates potential for improving automated analysis in nephropathology.