High-throughput image analysis with deep learning captures heterogeneity and spatial relationships after kidney...

Area of Science:

• Nephrology
• Computational Biology
• Pathology

Background:

• Acute kidney injury (AKI) recovery is highly variable, making analysis challenging.
• Spatial information from immunofluorescence is crucial but often limited by partial tissue analysis.
• Manual quantification of kidney injury is time-consuming and limits scalability.

Purpose of the Study:

• To develop and validate a deep learning approach for quantifying heterogeneous kidney injury responses.
• To enable analysis of larger tissue areas and sample sizes without specialized equipment or expertise.
• To track injury evolution and identify patterns of failed repair in kidney tissues.

Main Methods:

• Trained deep learning models on small datasets to identify stains and structures in kidney tissue.
• Applied the models to analyze folic acid-induced and ischemic AKI in mouse models.
• Quantified spatial heterogeneity in tubule repair and correlated failed repair with peritubular capillary density.

Main Results:

• Deep learning models achieved performance comparable to human observers in identifying tissue features.
• The approach accurately tracked AKI evolution, highlighting spatially clustered tubules that failed to repair.
• Failed repair post-ischemic injury was spatially correlated and inversely associated with capillary density.

Conclusions:

• Deep learning offers a versatile and accessible tool for quantifying spatially heterogeneous kidney injury.
• This method enhances the analysis of AKI recovery, revealing previously uncharacterized repair patterns.
• The findings provide insights into the mechanisms of failed repair and potential therapeutic targets.