Holistic fine-grained global glomerulosclerosis characterization: from detection to unbalanced classification.

Area of Science:

• Nephrology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Global glomerulosclerosis (GGS) has demonstrated diagnostic and prognostic value in kidney diseases like IgA nephropathy.
• Manual quantification of GGS subtypes is labor-intensive and resource-demanding.
• A need exists for automated methods to analyze GGS subtypes efficiently.

Purpose of the Study:

• To develop a fully automated pipeline for detecting and classifying global glomerulosclerosis (GGS) subtypes from whole slide images.
• To provide a quantitative analytical tool for fine-grained GGS characterization.
• To address challenges in unbalanced classification and integrate detection with classification.

Main Methods:

• A deep learning framework with a hierarchical two-stage design for GGS detection and classification.
• Incorporation of transfer learning techniques to enhance model generalizability and handle imbalanced datasets.
• Development of a WSI-to-results pipeline for efficient GGS characterization.

Main Results:

• The model achieved a 0.778-macro- $$ score for fine-grained GGS characterization when pretrained on a larger dataset.
• An AUC score of 0.994 was obtained for differentiating GGS from normal glomeruli on an external dataset.
• The developed algorithms demonstrated robustness to distribution shifts for glomeruli lesion classification.

Conclusions:

• The proposed methods significantly improve GGS detection and fine-grained classification performance.
• Both cross-validation and external validation confirmed the effectiveness of the approach.
• The open-source code and pretrained models are available to the research community.