Boundary-aware glomerulus segmentation: Toward one-to-many stain generalization

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Nephrology Research

Background:

• Whole-slide imaging (WSI) in nephropathology offers advanced diagnostic capabilities.
• Accurate segmentation of renal structures, particularly glomeruli, is crucial for diagnosis.
• Existing methods may struggle with diverse staining techniques and complex tissue structures.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) for accurate glomerulus segmentation in renal WSIs.
• To create an end-to-end network, DS-FNet, integrating semantic segmentation and boundary detection with an attention mechanism.
• To evaluate DS-FNet's performance across multiple datasets and staining methods.

Main Methods:

• Proposed a novel attention-aware convolutional neural network, DS-FNet, for glomerulus segmentation.
• Trained DS-FNet on periodic acid-Schiff (PAS)-stained WSIs.
• Validated DS-FNet on diverse datasets (HuBMAP, NEPTUNE, WSI_Fiocruz) with multiple staining techniques (PAS, PAMS, HE, Masson trichrome).
• Compared DS-FNet against six other deep learning models including U-Net variants and DeepLabV3+.

Main Results:

• DS-FNet achieved high performance, with a Dice Score (DSC) of 95.05% on the HuBMAP dataset.
• DS-FNet demonstrated superior average DSC on NEPTUNE and WSI_Fiocruz datasets compared to other networks.
• The network effectively segmented glomeruli across various staining methods, not limited to the training stain (PAS).

Conclusions:

• DS-FNet provides a robust and versatile solution for glomerulus segmentation in digital nephropathology.
• The proposed network shows consistent high performance across different datasets and staining techniques.
• This work represents a significant step towards automated analysis of renal WSIs, aiding pathologists in medical decision-making.