Abstract: PO0546
Magnetic Resonance Imaging-Based Renal Function Estimation Using a Machine Learning Approach
Session Information
- CKD Clinical, Outcomes, and Trials - 1
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: CKD (Non-Dialysis)
- 2102 CKD (Non-Dialysis): Clinical, Outcomes, and Trials
Authors
- Fukaya, Daichi, Department of Nephrology, Faculty of Medicine, Saitama Medical University, Iruma-gun, Saitama, Japan
- Inoue, Tsutomu, Department of Nephrology, Faculty of Medicine, Saitama Medical University, Iruma-gun, Saitama, Japan
- Kozawa, Eito, Department of Radiology, Faculty of Medicine, Saitama Medical University, Iruma-gun, Saitama, Japan
- Ishikawa, Masahiro, School of Clinical Engineering, Faculty of Health and Medical Care, Saitama Medical University, Hidaka-shi, Saitama, Japan
- Watanabe, Yusuke, Division of Dialysis Center and Department of Nephrology, Saitama Medical University International Medical Center, Hidaka-shi, Saitama, Japan
- Amano, Hiroaki, Department of Nephrology, Faculty of Medicine, Saitama Medical University, Iruma-gun, Saitama, Japan
- Kobayashi, Naoki, School of Clinical Engineering, Faculty of Health and Medical Care, Saitama Medical University, Hidaka-shi, Saitama, Japan
- Niitsu, Mamoru, Department of Radiology, Faculty of Medicine, Saitama Medical University, Iruma-gun, Saitama, Japan
- Okada, Hirokazu, Department of Nephrology, Faculty of Medicine, Saitama Medical University, Iruma-gun, Saitama, Japan
Background
In patients with deterioration of GFR with an unknown clinical course, it is quite difficult to determine whether the renal dysfunction is caused by a hemodynamic alteration or changes in the renal parenchyma, even when using kidney imaging. Therefore, to estimate renal function quantitatively based on the morphology of the renal parenchyma, we performed an advanced image analysis of renal magnetic resonance imaging (MRI) using machine learning (ML).
Methods
We used coronal DIXON water-dominant images obtained on a 3.0T MR device and a deep ML convolutional neural network (CNN) to evaluate renal function (eGFRcre). K-fold cross-validation (k = 5) was performed for the assessment of accuracy and generalization performance. The study protocol was approved by the IRB of our institute.
Results
A total of 196 patients (age, 57.9 ± 16.9 years; 128 males; CKD stage, G1 (n = 18), G2 (39), G3a (43), G3b (45), G4 (35), and G5 (16)) were included. After optimization of the CNN model, the accuracy, precision, recall, and f1-score of the confusion matrix, as well as the AUC of the ROC curve at thresholds of eGFRcre of 60, 45, and 30 were 0.80, 0.83, 0.90, 0.87, 0.86; 0.75, 0.71, 0.84, 0.77, 0.83; and 0.76, 0.80, 0.90. 0.85, 0.83, respectively. The output value of the CNN also showed a significant positive correlation with the normalized eGFRcre of the subjects (R2 = 0.46, P < 0.01). When the difference in signal intensity between the renal cortex and medulla, as measured based on the region of interest method, was used as a diagnostic index, the accuracy was the same as that of ML if the threshold was eGFRcre 30 (AUC of the ROC curve, 0.84). Conversely, when the threshold was set at eGFRcre 45 or 60, the accuracy deteriorated gradually (AUC, 0.80 and 0.73, respectively).
Conclusion
Compared with the classical method, in which only the signal intensity is used, the ML approach was able to quantitatively evaluate differences in renal morphology regarding a wide range of renal functions. Our results may have clinical applications for assessing the cause of changes in kidney function in the conditions in which renal function and morphology diverge, e.g., in the early stages of acute kidney injury, renovascular hypertension, and therapeutic interventions that cause hemodynamic alterations.
Funding
- Government Support - Non-U.S.