Up to 60% of adults and 80% of children with systemic lupus erythematosus (SLE) develop nephritis (LN), with 10?30% progressing to end-stage renal disease (ESRD). The gold standard for diagnosis of LN is a renal biopsy. Histological parameters remain the best predictors of ESRD. Despite being the gold standard, histological diagnosis of LN has several shortcomings. In multiple inter-observer renal pathology assessment studies reported thus far, the inter- pathologist correlation coefficients, or concordance, in assessing most histological parameters have been sub-optimal. This has provided the impetus for the current proposal. We propose to leverage the power of computer vision and deep learning to build a classifier that rivals the best-trained renal pathologists in making a histological diagnosis of LN using current diagnostic criteria. We propose to train a deep convolutional neural network to distinguish the different LN classes, and to identify a full spectrum of histological attributes useful for diagnosis. We will compare the performance of the newly generated neural network in scoring glomerular/tubulo-interstitial features and LN classes, against a panel of human renal pathologists. Finally, we propose to build a neural network that can predict clinical outcome based on baseline renal pathology. Reliable and reproducible classification of LN could dramatically improve patient management and long-term renal and patient survival.
Despite being the gold standard, histological diagnosis of lupus nephritis is imprecise, and marked by significant inter-pathologist discordance in readings. We propose to leverage the power of computer vision and deep learning to build a classifier that rivals the best-trained renal pathologists in making a histological diagnosis of lupus nephritis. Reliable and reproducible classification of LN could dramatically improve patient management and long-term renal and patient survival.
