Automated detection algorithm for C4d immunostaining showed comparable diagnostic performance to pathologists in...

Area of Science:

• Nephrology
• Pathology
• Artificial Intelligence in Medicine

Background:

• Deep learning-based image analysis offers potential for enhanced diagnostic accuracy and efficiency in pathology.
• C4d immunostaining in renal allografts is crucial for diagnosing antibody-mediated rejection.
• Variability in human interpretation necessitates objective and reproducible diagnostic tools.

Purpose of the Study:

• To assess the diagnostic performance of a proposed deep learning-based C4d detection algorithm in renal allografts.
• To compare the algorithm's diagnostic accuracy against experienced pathologists.
• To analyze the association of the algorithm's results with clinical data and patient outcomes.

Main Methods:

• A deep learning algorithm was developed for C4d detection in renal allograft immunostaining slides.
• Slides from two institutions were evaluated independently by three pathologists and the algorithm using Banff 2017 criteria.
• Algorithm performance was compared to individual pathologists and a consensus diagnosis; clinicopathological associations were analyzed.

Main Results:

• Pathologist reproducibility was fair to moderate (kappa 0.36-0.54), similar to algorithm-pathologist agreement (kappa 0.34-0.51).
• The algorithm achieved substantial concordance with the consensus diagnosis (kappa = 0.61).
• Algorithm-predicted C4d scores significantly correlated with microvascular inflammation, donor-specific antibody detection, and shorter graft survival.

Conclusions:

• The deep learning-based C4d detection algorithm demonstrates diagnostic performance comparable to human pathologists.
• The algorithm shows significant associations with key indicators of renal allograft dysfunction and poor outcomes.
• This AI tool holds promise for improving the objectivity and efficiency of C4d assessment in renal transplantation.