Deep Learning Morphometric Analysis on Protocol Biopsies Predicts Future Graft Function.

Area of Science:

• Nephrology
• Transplant Medicine
• Computational Pathology

Background:

• The Banff classification has limitations in predicting outcomes from protocol biopsies without specific lesions.
• Traditional morphometry offers precise data but is time-consuming, hindering clinical application.

Purpose of the Study:

• To evaluate the efficacy of automated morphometric analysis combined with deep learning and machine learning in predicting long-term kidney transplant function.
• To assess the predictive value of these computational methods for estimated glomerular filtration rate (eGFR) at three years post-transplant.

Main Methods:

• A retrospective study involving 367 kidney transplant recipients with protocol biopsies.
• Eight deep learning algorithms extracted 23 morphometric parameters from whole-slide images (WSI).
• Ten machine learning models were trained and validated to predict 3-year eGFR.

Main Results:

• The best performing models (Kernel Ridge and Bayesian) achieved a mean absolute error (MAE) of 11 ± 1 ml/min/1.73 m² in the training set.
• External validation demonstrated a good association between predicted and observed eGFR (MAE = 13 ± 11 ml/min/1.73 m², r = 0.68).
• No significant difference was found between predicted and observed eGFRs after bias correction.

Conclusions:

• Automated morphometric analysis integrated with machine learning models shows promise for predicting kidney transplant outcomes.
• This computational approach can aid in forecasting long-term graft function, specifically eGFR at three years post-biopsy.