Abstract: SA-PO048
Contribution of Glomerular Phenotype to Digital Classification of Diabetic Nephropathy
Session Information
- Engineering-Based Approaches to Problems in Nephrology
November 09, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Bioengineering
- 300 Bioengineering
Authors
- Ginley, Brandon, SUNY Buffalo, Buffalo, New York, United States
- Lutnick, Brendon, SUNY Buffalo, Buffalo, New York, United States
- Jen, Kuang-Yu, University of California, Davis, Sacramento, California, United States
- Fogo, Agnes B., Vanderbilt University Medical Center, Nashville, Tennessee, United States
- Jain, Sanjay, Washington University School of Medicine, St. Louis, Missouri, United States
- Rosenberg, Avi Z., Johns Hopkins University, Baltimore, Maryland, United States
- Walavalkar, Vighnesh, UC San Francisco Medical Center, San Francisco, California, United States
- Tomaszewski, John E., SUNY Buffalo, Buffalo, New York, United States
- Rossi, Giovanni maria, Johns Hopkins University, Baltimore, Maryland, United States
- Sarder, Pinaki, SUNY Buffalo, Buffalo, New York, United States
Background
Diabetic nephropathy (DN) is a leading cause of kidney disease; renal pathologists assess its pathology via visual interpretation of biopsied tissue in the form of a digitized whole slide image (WSI). Inter-rater agreement increases with consensus classifications like the Tervaert approach, but reproducibility could still be improved. Computation can unify interpretation of image structure. We engineered a complete start-to-finish glomerular detection and classification pipeline for digitized biopsies of DN. Further, we investigated the glomerular phenotypes and features that it relies on to make decisions.
Methods
We studied 54 patients. Glomeruli were detected from WSIs using our previously published method for WSI segmentation. Glomerular structure was condensed to a three-component system that facilitates detection in widely varying phenotypes. Handcrafted features (n = 232) were used to quantify glomerular structures. Glomerular features from a single biopsy were fed as a sequence to a recurrent neural network (RNN) which yields a continuous number representing Tervaert class. Glomeruli and features were dropped from the network one-by-one; the resultant shift in predicted class was used as a proxy to investigate how much each glomerulus and feature contributed to the overall output. We trained our method by taking one renal pathologist as ground truth, and compared its performance against two other renal pathologists.
Results
Digital classification agreed with the ground truth renal pathologist with linear weighted Cohen kappa κ = 0.55, 95% confidence interval (CI) [0.5, 0.6]. The other two other renal pathologists agreed with the first with κ1 = 0.48, 95% CI [0.32, 0.64], and κ2 = 0.68, [0.5, 0.86]. The features that hurt network performance the most when dropped out included features quantifying color of glomerular structures as well as those quantifying mesangial width. The network was most sensitive to dropped glomeruli when there were fewer total glomeruli contained in the biopsy.
Conclusion
Our digital classification reaches agreement similar to renal pathologists and makes decisions intuitively. Digital quantification of renal tissue can enhance clinical workflow by improving precision used to describe disease state.
Funding
- NIDDK Support