A Renal Revolution

Kidney disease diagnosis is not always easy – renal pathology is an uncommon specialty and reading kidney biopsies can present a challenge. But what if computational pathology could help improve the diagnostic process and the accuracy of the ultimate diagnosis? Two recent studies have taken on the task of building a better mousetrap by creating computational tools to tackle areas of difficulty in kidney disease diagnosis.

The first study, performed by researchers at the University at Buffalo, combined image analysis and machine learning into a digital algorithm to classify renal biopsies from patients with diabetic nephropathy. Although glomerular structure is complex and can be difficult for even human pathologists to fully quantify, the researchers provided a set of simplified components for the algorithm to use in its classification. Ultimately, the digital classifications of biopsies from 54 diabetic nephropathy patients substantially agreed with those of three different human pathologists – and the algorithm was able to detect “glomerular boundaries […] with 0.93±0.04 balanced accuracy, glomerular nuclei with 0.94 sensitivity and 0.93 specificity, and glomerular structural components with 0.95 sensitivity and 0.99 specificity (1).”

The second study, conducted at Radboud University Medical Center, used a convolutional neural network to extend the analysis to multiple tissue classes in kidney transplant biopsies – an area where work is already time-consuming and unreliable (2). The network’s performance varied depending on the tissue type; glomeruli were a particular strength, with the network detecting 92.7 percent of all glomeruli and exhibiting a 10.4 percent false positive rate (3). Although there remains room for improvement, this first convolutional neural network of its type heralds future possibilities for deep learning and neural networks in the day-to-day diagnostic workflow.