Abstract: FR-PO301
Evaluation of an Artificial Deep Neural Network for Fully Automated Segmentation of Individual Kidneys and Liver in Patients with Polycystic Kidney Disease
Session Information
- Cystic Kidney Diseases - II
November 03, 2017 | Location: Hall H, Morial Convention Center
Abstract Time: 10:00 AM - 10:00 AM
Category: Genetic Diseases of the Kidney
- 801 Cystic Kidney Diseases
Authors
- van Gastel, Maatje D.A., UMC Groningen, Groningen, Netherlands
- Edwards, Marie E., Mayo Clinic, Rochester, Minnesota, United States
- Torres, Vicente E., Mayo Clinic, Rochester, Minnesota, United States
- Gansevoort, Ron T., UMC Groningen, Groningen, Netherlands
- Kline, Timothy L., Mayo Clinic, Rochester, Minnesota, United States
Background
Polycystic kidney disease (PKD) leads to cyst formation in kidneys and often liver, with marked increase in total kidney volume (TKV) and total liver volume (TLV). TKV is recognized by the FDA and EMA as a biomarker that plays an important role in risk prediction in PKD. Thus far, there is no optimal alternative for the laborious and expensive measurement of manually tracing kidneys and liver in radiological imaging examinations. Therefore, we developed and validated a fully automated segmentation method for TKV and TLV measurement. This is the first study to propose a solution for measuring individual kidney volumes, and liver automatically in radiological examinations of patients with PKD.
Methods
The automated approach that was developed was a deep learning network optimized to perform voxel-based classification. The network was first trained (80%) and validated (20%) on a set of 100 abdominal MRIs (T2 weighted HASTE or TRUFI coronal sequences) of patients with PKD which had both kidneys and liver segmented manually. A test set of 60 patients was used to evaluate the performance of the developed automated method.
Results
TKV as well as TLV measured using the deep neural network correlated highly with manually traced TKV and TLV (ICC 0.996 and 0.994, resp.), with a bias and precision of -0.3%±4.6% for TKV and 2.5%±4.9% for TLV with maximal percentage differences observed being 8.5% and 11.2%, respectively. No proportional bias was observed, meaning that percentage differences between both methods are regardless of kidney or liver size.
Conclusion
This is the first fully automated segmentation method that measures individual kidney volumes, TKV and TLV almost as accurate as manual tracing, that has an inter-reader variability of 2.3%. The developed technique will facilitate future studies where automated and reproducible measurement of individual kidney volumes, TKV, and TLV are needed to assess (i) disease severity, (ii) progression of the disease, and (iii) treatment response.
Funding
- Other NIH Support