Deep Learning Morphometric Analysis on Protocol Biopsies Predicts Future Graft Function

<h2>Abstract</h2><h3>Introduction</h3> The predictive value of Banff Classification in protocol transplant biopsies without specific lesions is limited. Morphometry provides precise data on microstructures, surpassing semi-quantitative scores but is time-consuming. This study evaluates whether automated morphometric analysis with deep learning can predict glomerular filtration rate at three years using machine learning. <h3>Methods</h3> This retrospective study included kidney transplant recipients who underwent protocol biopsy without specific lesion. The models were trained and tested on the Training/Test cohort, with external validation on the Application cohort. Eight deep learning algorithms extracted 23 morphometric parameters from whole-slide images. Ten machine learning models were tested for three years glomerular filtration rates prediction. <h3>Results</h3> A total of 367 patients were included. The means three-year estimated glomerular filtration rates were 53±23 and 53±22 mL/min/1.73m² in the Training/Test and Application cohorts, respectively. In the Training/Test cohort, estimated glomerular filtration rate correlated negatively with interstitial fibrosis (r=-0.33; p<0.001), tubular atrophy (r=-0.39; p<0.001), artery luminal stenosis (r=-0.29; p<0.001), and positively with glomerular density (r=0.16; p<0.05) and glomerular epithelial (r=0.33; p<0.001), endothelial (r=0.30; p<0.001), and mesangial (r=0.25; p=0.002) cell densities. Kernel Ridge and Bayesian models achieved the best predictions (Mean Absolute Error=11±1 mL/min/1.73m²). External validation showed good association between predicted with Kernel Ridge model and observed estimated glomerular filtration rate (Mean Absolute Error=13±11 mL/min/1.73m², r=0.68; p<0.001). After correction with Bland Altman Bias, paired analysis showed no significant difference between predicted and observed estimated glomerular filtration rates (p=0.953). <h3>Conclusion</h3> Integrating Automated morphometric analyses into machine learning models may help predict glomerular filtration rates three years after protocol biopsies.