Abstract: TH-OR066
Predicting APOL1 Risk Category from Kidney Donor Biopsies Using Deep Learning
Session Information
- Harnessing Molecular, Machine-Learning, and Genomic Innovations in Pathology
October 25, 2018 | Location: 24A, San Diego Convention Center
Abstract Time: 05:30 PM - 05:42 PM
Category: Pathology and Lab Medicine
- 1502 Pathology and Lab Medicine: Clinical
Authors
- Xu, Jun, Nanjing University of Information Science and Technology, Nanjing, China
- Janowczyk, Andrew, Case Western Reserve University, Cleveland, Ohio, United States
- Barisoni, L., U. Miami, Miller School of Medicine, Miami Beach, Florida, United States
- Cai, Chengfei, Nanjing University of Information Science and Technology, Nanjing, China
- Nirschl, Jeffrey J., University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Palmer, Matthew, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Feldman, Michael, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Chen, Dhruti P., University of North Carolina, Chapel Hill, North Carolina, United States
- O'Toole, John F., Cleveland Clinic, Cleveland, Ohio, United States
- Zaky, Ziad S., Cleveland Clinic, Cleveland, Ohio, United States
- Poggio, Emilio D., Cleveland Clinic, Cleveland, Ohio, United States
- Sedor, John R., Cleveland Clinic, Cleveland, Ohio, United States
- Madabhushi, Anant, Case Western Reserve University, Cleveland, Ohio, United States
Background
African Americans (AA) subjects with two APOL1 gene variants (G1 and G2) are at higher risk (HR) to develop non-diabetic kidney disease, while individuals with 0 or 1 APOL1 variants are at low risk (LR). Conventional visual histologic analysis of renal biopsies does not identify structural differences between LR and HR subjects with or without kidney disease. In this work we sought to evaluate whether deep-learning (DL) based analysis could help identify phenotypic representations that were predictive of APOL1 variants from whole slide images (WSI) of kidney biopsies.
Methods
WSI’s from kidney allograft implant biopsies stained with H&E from 19 APOL1 LR and 12 APOL1 HR AA donors were annotated by a pathologist to identify artifact-free regions of interest (ROIs). The ROIs from the training set (8 LR and 5 HR) were divided into non-overlapping 256x256 patches for processing at 40X magnification. Each patch was assigned a HR or LR label based on the subject’s APOL1 genotype. An 8 layer AlexNet model was trained to predict the likelihood that a previously unseen test input patch belongs to the APOL1 HR genotype. This patch-wise output was then overlaid as a heatmap on the original image such that red and blue pixels predicts HR and LR class, respectively. Majority voting across the individual patches were aggregated to generated the combined prediction of APOL1 risk status as either HR or LR. (Figure 1).
Results
The prediction accuracy of trained model was 84.67% and 91.02% for LR and HR cases respectively (n=18, LR=11, HR=7) on the test set.
Conclusion
In this preliminary study, the DL model was able to distinguish between APOL1 HR and LR genotypes in AA patients. Future work will involve larger scale multi-site independent validation.
Figure 1: Flowchart illustrating the training and testing for predicting APOL1 risk status, where the bottom right image heatmap is the output from the DL classifier where redder pixels indicate higher likelihood of belonging to the HR class.
Funding
- Other NIH Support