Abstract: PO0530
Automation of Renal Blood Flow Analysis from Dynamic Phase-Contrast MRI with Deep Learning
Session Information
- CKD Health Services Research
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2101 CKD (Non-Dialysis): Epidemiology, Risk Factors, and Prevention
Authors
- Martínez Mora, Andrés, Antaros Medical, Mölndal, Sweden
- Lindgren, Dag, Antaros Medical, Mölndal, Sweden
- Englund, Camilla, Antaros Medical, Mölndal, Sweden
- Johansson, Lars, Antaros Medical, Mölndal, Sweden
- Hockings, Paul, Antaros Medical, Mölndal, Sweden
- Kullberg, Joel, Antaros Medical, Mölndal, Sweden
Background
Phase-contrast magnetic resonance imaging (PC-MRI) allows to assess renal blood flow (RBF), an important parameter in the development of chronic kidney disease (CKD). RBF assessments require time-consuming and observer-dependent delineations of the renal arteries. Thus, we have developed and evaluated a fully-automated deep learning model for renal artery segmentation.
Methods
PC-MRI data came from 131 subjects, four studies, three MRI vendors and a range of velocity encodings. The deep learning model (DL) was a deeply-supervised attention U-Net with residuals with the result re-introduced in a second iteration. Flow was estimated by integrating the flow values in the segmentations. Segmentation and flow results were compared for cross-validation (CV, 73 subjects) against manual delineations and reference flow measurements from external software. The remaining data (Extra) only had reference flow measurements, being only evaluated for flow.
Results
In 4-fold CV, a segmentation accuracy of Dice 0.71 ± 0.21 was obtained. Although most segmentations were relatively accurate, the model failed in ten out of 144 arteries. Flow measurements were relatively highly correlated in CV with no significant deviation from the reference: (r=0.84, DL: 5.8±3.0 ml/s vs Ref: 5.8±3.0 ml/s, p=0.98). The Extra set provided a high correlation and no significant deviation (r=0.94, DL: 6.4±2.8 ml/s vs Ref: 6.7±2.9 ml/s, p=0.11).
Conclusion
The method shows promise to support RBF measurements from PC-MRI. It may save analysis time and increase objectivity in the future. More high quality and representative training data are likely to improve accuracy and generalizability.
Best segmentation results (a,b) and worst (c,d). Green shows true positives, blue false negatives, and red false positives. (e) and (f) show the linear regression plots comparing DL vs Ref for CV and Extra sets, respectively, with the fitting line in thick blue and the ideal line in thin gray
Funding
- Government Support - Non-U.S.