Deep learning-based histopathological assessment of tubulo-interstitial injury in chronic kidney diseases

Chronic kidney disease (CKD) causes progressive and irreversible damage to the kidneys. Renal biopsies are essential for diagnosing the etiology and prognosis of CKD, while accurate quantification of tubulo-interstitial injuries from whole slide images (WSIs) of renal biopsy specimens is challenging with visual inspection alone.

We develop a deep learning-based method named DLRS to quantify interstitial fibrosis and inflammatory cell infiltration as tubulo-interstitial injury scores, from WSIs of renal biopsy specimens. DLRS segments WSIs into non-tissue areas, glomeruli, tubules, interstitium, and arteries, and detects interstitial nuclei. It then quantifies these tubulo-interstitial injury scores using the segmented tissues and detected nuclei.

Applied to WSIs from 71 Japanese CKD patients with diabetic nephropathy or benign nephrosclerosis, DLRS-derived scores show concordance with nephrologists' evaluations. Notably, the DLRS-derived fibrosis score has a higher correlation with the estimated glomerular filtration rate (eGFR) at biopsy than scores from nephrologists' evaluations. Validated on WSIs from 28 Japanese tubulointerstitial nephritis patients and 49 European-ancestry patients with nephrosclerosis, DLRS-derived scores show a significant correlation with eGFR. In an expanded analysis of 238 Japanese CKD patients, including 167 from another hospital, deviations in eGFR from expected values based on DLRS-derived scores correlate with annual eGFR decline after biopsy. Inclusion of these deviations and DLRS-derived fibrosis scores improve predictions of the annual eGFR decline.

DLRS-derived tubulo-interstitial injury scores are concordant with nephrologists' evaluations and correlated with eGFR across different populations and institutions. The effectiveness of DLRS-derived scores for predicting annual eGFR decline highlights the potential of DLRS as a predictor of renal prognosis.