Streamlining the Histopathologic Workflow in Diabetic Kidney Disease with Artificial Intelligence

Area of Science:

• Nephrology
• Computational Pathology
• Artificial Intelligence

Background:

• Diabetic kidney disease (DKD) pathology assessment in animal models is slow and subjective.
• Limited human kidney biopsy data impedes translational research from animal models to humans.

Purpose of the Study:

• To develop an AI-driven workflow for efficient and accurate histopathological assessment in diabetic nephropathy animal models.
• To enhance the translation of AI scoring models from preclinical animal studies to human kidney biopsies.

Main Methods:

• Developed an AI workflow including glomeruli detection, an expert annotation tool for fast scoring, and automated scoring.
• Utilized self-supervised learning on unlabeled preclinical data to improve AI performance and reduce bias.
• Implemented a method to adapt AI models to human kidney features for cross-species translation.

Main Results:

• AI annotation tool reduced scoring time by 80%; automated scoring reduced it by 90%, outperforming expert agreement.
• Self-supervised learning achieved a κ of 0.78; combined with supervised learning, performance improved to κ = 0.84, reducing bias.
• The translational approach achieved a κ of 0.63 on human glomeruli, reducing the translational gap by 45%.

Conclusions:

• AI-assisted scoring accelerates pathology readouts and reduces pathologist workload in pharmaceutical settings.
• Self-supervised learning captures kidney morphology, minimizes bias, and bridges the translational gap in DKD drug development.