Deep Learning Global Glomerulosclerosis in Transplant Kidney Frozen Sections

Area of Science:

• Nephrology
• Digital Pathology
• Artificial Intelligence

Background:

• Limited supply of transplantable kidneys necessitates accurate viability assessment to minimize organ discard.
• Pathologist evaluation of donor kidney biopsies, specifically percent global glomerulosclerosis, is critical for transplant eligibility but suffers from significant inter- and intra-observer variability.
• Automated methods for glomerulosclerosis assessment could enhance evaluation consistency, throughput, and reduce pathologist workload.

Purpose of the Study:

• To develop and validate a deep learning model for automated identification and classification of normal and sclerotic glomeruli in whole-slide images of donor kidney frozen section biopsies.
• To assess the model's performance against expert renal pathologists and evaluate its robustness to frozen section artifacts.
• To overcome technical challenges in applying pre-trained convolutional neural networks (CNNs) to whole-slide image analysis for kidney biopsy evaluation.

Main Methods:

• Development of a deep learning model, extending a pre-trained CNN, to analyze whole-slide images of donor kidney frozen sections.
• Training the model on 48 whole-slide images for slide-level evaluation of glomeruli (normal vs. sclerotic).
• Comparison of the whole-slide image model's performance against a patch-based model and expert renal pathologists, considering accuracy and speed.

Main Results:

• The developed deep learning model achieved slide-level evaluation performance comparable to expert renal pathologists using only 48 whole-slide images.
• The model demonstrated robustness to slide preparation artifacts common in frozen sections.
• The whole-slide image analysis model significantly outperformed a patch-based model in both accuracy and speed for glomeruli classification.

Conclusions:

• The developed deep learning model offers a reliable, automated method for assessing percent global glomerulosclerosis in donor kidney biopsies.
• This AI-driven approach has the potential to decrease evaluation variability, increase efficiency, and aid in donor kidney selection for transplantation.
• As the first model for classifying glomeruli in frozen kidney sections, it represents a significant advancement for kidney transplantation and donor organ evaluation.