Kidney Week

Abstract: TH-PO1039

Machine Learning and Multiparametric MRI for Noninvasive Diagnosis of the Etiology of CKD

Session Information

• CKD Progression and Complications: Diagnosis, Prognosis, Risk Factors
  November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Augmented Intelligence, Digital Health, and Data Science

• 300 Augmented Intelligence, Digital Health, and Data Science

Authors

• Inoue, Tsutomu, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Tsutomu Inoue,

• Kozawa, Eito, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Eito Kozawa,

• Ishikawa, Masahiro, Kinki Daigaku, Higashiosaka, Osaka, Japan

Masahiro Ishikawa,

• Amano, Hiroaki, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Hiroaki Amano,

• Kosakai, Wakako, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Wakako Kosakai,

• Watanabe, Yusuke, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Yusuke Watanabe,

• Tomori, Koji, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Koji Tomori,

• Kobayashi, Naoki, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Naoki Kobayashi,

• Okada, Hirokazu, Saitama Ika Daigaku, Iruma-gun, Saitama, Japan

Hirokazu Okada,

Background

Multiparametric magnetic resonance imaging (MRI) has the potential to provide various types of biological information about the kidney. In this study, we aimed to diagnose the underlying etiology of chronic kidney disease (CKD) non-invasively by combining MRI images and machine learning.

Methods

T1-weighted images (water-weighted images using the Dixon method), T1 value maps, T2* value maps of blood oxygen level-dependent MRI, perfusion maps of arterial spin-labeling, fractional anisotropy value maps from diffusion tensor imaging, and apparent diffusion coefficient value maps from diffusion-weighted imaging were used. We calculated the cortical values and cortical-medullary gradients using a 12-layer concentric object method. We created a multiclass classifier using a support vector machine and features such as MRI measurement values, age, estimated glomerular filtration rate (eGFR) at the time of imaging, and hemoglobin value. K-fold cross-validation was used to evaluate classifier accuracy.

Results

A total of 197 patients [60.9 ± 14.9 years old, 65.0% male, 15.2% with diabetic kidney disease (DKD), 34.0% with chronic glomerulonephritis (CGN), 50.8% with nephrosclerosis (NS), and a mean eGFR of 42.3 ± 22.2 ml/min/1.73 m²] were included. After model optimization, we obtained a relatively good overall accuracy of 0.65, and area under the curve values of 0.72 for DKD, 0.76 for CGN, and 0.73 for NS on the receiver operating characteristic curve. There were some estimation errors, particularly in cases diagnosed with DKD by physicians, and accuracy of the estimation tended to be low. Among cases of DKD, there were also cases that showed imaging characteristics similar to those of NS or CGN, suggesting an overlap of pathologies.

Conclusion

Conventional and functional kidney MRI and machine learning/augmented intelligence diagnosis have the potential to be useful noninvasive diagnostic tools, and further improvements in accuracy can be achieved by refining the features. We observed some cases diagnosed with DKD by nephrologists based on the clinical course and results of blood/urine test that showed MR imaging characteristics resembling those of NS or CGN. Kidney biopsy or renal MRI may be effective for accurate diagnosis and determination of an appropriate treatment plan.