# Texture Analysis of Kidney MRI Using Diffusion-Weighted Imaging and Intravoxel Incoherent Motion for Classifying the Severity of Chronic Kidney Diseases

## TL;DR

This study explores using MRI texture analysis with diffusion-weighted imaging to classify the severity of chronic kidney disease, finding that ADC-based models perform best.

## Contribution

The study demonstrates that texture analysis of ADC maps outperforms IVIM-derived parameters in classifying CKD severity.

## Key findings

- ADC map-based QDA models achieved the highest AUC of 0.851 for CKD severity classification.
- Texture features from the left kidney showed slightly better results than those from the right kidney.
- IVIM-derived parameters did not outperform ADC-based models in classifying renal dysfunction.

## Abstract

Introduction

Chronic kidney disease (CKD) is a significant global health concern, and noninvasive imaging biomarkers for chronic kidney disease have been investigated using magnetic resonance imaging (MRI), including diffusion‑weighted imaging (DWI), blood oxygenation level-dependent (BOLD) imaging, arterial spin labeling (ASL), and T1 mapping; however, there are currently no widely accepted imaging biomarkers for the non-invasive assessment of renal dysfunction severity. DWI and intravoxel incoherent motion (IVIM), when combined with texture analysis (TA), provide a promising approach for the non-invasive assessment of renal microstructure. This study investigated the utility of TA applied to DWI/IVIM-derived maps for assessing the severity of renal dysfunction.

Materials and methods

We retrospectively analyzed kidney MRI data from 68 patients with CKD who underwent DWI-IVIM. Data were categorized into three groups based on the estimated glomerular filtration rates (eGFRs). Two-dimensional segmentation was performed on the apparent diffusion coefficient (ADC), true diffusion coefficient, pseudo-perfusion diffusion coefficient, and perfusion fraction map. After extracting 93 texture features (TFs), we employed a sequential feature selection algorithm for feature selection. Multiclass classification models were developed using linear discriminant analysis, quadratic discriminant analysis (QDA), support vector machines, k-nearest neighbors, decision trees, and random forest classifiers. To evaluate model performance, we performed a 10‑fold cross‑validation 100 times and used the mean of the resulting performance measures as the performance estimate.

Results

TA models demonstrated acceptable discriminatory performance, with the highest performance achieved by the ADC map-based QDA model (area under the curve or AUC 0.851±0.010). The IVIM-derived parameters did not outperform the ADC-based models. TFs derived from the left kidney showed slightly better results than those derived from the right kidney. There were no significant differences between TFs derived from the renal medulla and the cortex.

Conclusion

Although the combination of DWI-IVIM and TA has the potential to evaluate renal impairment in patients with CKD, the TA model based on the ADC map achieved higher accuracy than the models based on other IVIM parameters. Consequently, the utility of TA using ADC maps was reaffirmed. Future research with larger cohorts and improved image quality is needed to further refine the non-invasive assessment of renal dysfunction.