Digital pathology for kidney transplant precision diagnostics – DEEPGRAFT

Rejection currently represents the major cause of allograft failure worldwide, with immediate consequences for the patients in terms of mortality, morbidity and cost for the society. Improvement of diagnostics and prognostics are required for better patient stratification, personalized treatment, and improved allograft survival.
The current management of the transplanted patient is based on a non-standardized approach; currently available follow-up markers are left to the subjective assessment of the clinician or pathologist, who cannot accurately assess them both qualitatively and quantitatively 1) in an integrative manner (the whole available markers) 2) contextually (compared to a reference cohort with the same characteristics). In addition, current approaches to control allograft damage use non-specific and non-sensible markers when examined individually (creatininemia, proteinuria etc.), and do not provide information on the underlying etiopathology of the rejection.
In kidney transplant, the current gold standard for rejection diagnosis relies on the Banff classification, which is based on the histological grading of several lesions reflecting acute and chronic tissue damage and inflammation status. Despite its usefulness, several limitations have been pointed out including the low reproducibility and the use of semi-quantitative scores to quantify continuous changes, leading to a potential loss of diagnostic resolution, possible misclassification leading to suboptimal therapeutic choice and consequences for allograft and patient survival.
We hypothesized that a digital pathology approach could lead to an automated diagnostic, improving its accuracy and reproducibility, together with optimizing patient risk stratification. This project will set up innovative tools for precision diagnostics of rejection in kidney transplantation by creating Digital Pathology algorithms as companions to classical histology, integrated into a probabilistic system of precision medicine. To achieve this goal, we will rely on a prospective and well characterized multicentric prospective cohort of kidney transplant recipient (n=6,000) with extensive phenotype. We will digitalize and annotate associated allograft biopsies (n=10,000) (WP1), apply artificial intelligence and deep learning methods to automate and improve histological diagnostics (WP2), and integrate these digital pathology diagnostics into integrative prognostic models aiming to improve graft loss prediction, leading to better patient stratification (WP3). Long term goal for the diagnostic tools from WP2 will be to be provided as clinical decision support systems to assist physicians in their everyday work, resulting in better patient care.