Abstract: TH-PO828
A Convolutional Neural Network for Large-Scale Segmentation of Kidneys in Autosomal Polycystic Kidney Disease
Session Information
- Cystic Kidney Diseases: Clinical
November 07, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Genetic Diseases of the Kidneys
- 1001 Genetic Diseases of the Kidneys: Cystic
Authors
- Antiga, Luca, Orobix Srl, Bergamo, Italy
- Mohiuddin, Dr. imtiaz H., Otsuka Pharmaceutical, Rockville, Maryland, United States
- Dixon, Mary, Innovenn, Madison, Wisconsin, United States
- Lutz, Annie, Innovenn, Inc., Madison, Wisconsin, United States
Background
Total kidney volume (TKV), along with age and Glomerular Filtration Rate (eGFR), is an early prognostic marker of progression in autosomal dominant polycystic kidney disease (ADPKD). Current manual or semi-automated methods for estimation of TKV from imaging data are laborious, time-consuming approximations subject to human perception and experience; this has hampered a widespread adoption of TKV as a biomarker in ADPKD. We report the development and performance of a fully automated method for kidney segmentation and TKV estimation from magnetic imaging (MR) data in patients with ADPKD on a large patient cohort using a deep learning approach. In addition, we describe how such an estimate can be employed for predicting disease progression and monitoring progression, with the aim of supporting clinical management.
Methods
We employ a fully-convolutional neural network based on the volumetric U-net architecture, trained on an extensive dataset of 1620 T2-weighted magnetic resonance imaging scans extracted from the multicenter TEMPO3:4 trial (NCT00428948); expert outlines were available as ground truth. The method is validated on 490 scans, not included in the training dataset, extracted from 179 individual subjects. Based on the data from the same trial, we develop a similarity model for the prediction of the expected TKV growth over time.
Results
We obtained a 90th percentile estimation error of TKV and its change over time of 13% and 11% of the baseline volume, respectively. We predict 3-year TKV based on baseline characteristics with R2 of 0.954 on the TEMPO3:4 placebo data.
Conclusion
The present work represents the first, large-scale example of fully automated TKV estimation in ADPKD that has been trained and validated on a large-scale, multi-centric dataset. When coupled with clinical data from the same trial, we demonstrate the ability of a machine learning algorithm to predict likely TKV progression with high accuracy. This prognostic information combined with other clinical findings may support clinical care.