Abstract: SA-PO046
Unsupervised Modeling of Glomeruli for Diabetic Nephropathy Staging in Renal Biopsies
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
- Lutnick, Brendon, SUNY Buffalo, Buffalo, New York, United States
- Ginley, Brandon, SUNY Buffalo, Buffalo, New York, United States
- Jen, Kuang-Yu, University of California, Davis, Sacramento, California, United States
- Jain, Sanjay, Washington University School of Medicine, St. Louis, Missouri, United States
- Sarder, Pinaki, SUNY Buffalo, Buffalo, New York, United States
Background
As biological science pushes for computational analysis, the success of deep learning has obliged its adoption. However, biological datasets are different from the well-annotated standardized data used to develop such algorithms. Due to availability of unlabeled biological data, we have tested a variational autoencoder (VAE) for unsupervised modeling of glomeruli images (without labels). We show that encoded features allow interpolation between image states and are predictive of biopsy-level Tervaert classing of diabetic nephropathy (DN).
Methods
A VAE was trained using 87K 256x256 PAS and H&E stained glomeruli images, segmented from whole-slide kidney biopsies. The VAE encodes the images into a code of 200 numbers able to be decoded back to the input image. This technique automatically clusters similar images together in the code space. To show the relevance of our trained VAE, we model Tervaert DN class for expert staged human biopsies using the image codes as sequential input for a recurrent neural network (RNN). Namely a network model which predicts DN class from sequencial reading of glomeruli codes, similar to how experts read biopsies.
Results
While our decoded (simulated) glomeruli images lack detail, exploration of the image code space shows smooth interpolation between holdout images, verifying the continuity of the encoded data distribution. Figure 1 shows simulated samples between class I and IV DN glomeruli generated by our VAE. Despite having only 54 staged biopsies, an RNN model using our image codes for biopsy level DN class prediction predicts <1 class off with a mean square error of 0.971 and linear weighted Cohen’s kappa of 0.402 with 10-fold cross validation.
Conclusion
Further refinement of the VAE architecture is needed to produce sharp images, but current image codes are predictive of Tervaert DN class using supervised regression modeling. In the future, this promises powerful ways to incorporate unlabeled biological data to augment machine learning training sets of medical images.
Funding
- NIDDK Support