Three-dimensional convolutional neural network-based classification of chronic kidney disease severity using kidney MRI

Abstract

A three-dimensional convolutional neural network model was developed to classify the severity of chronic kidney disease (CKD) using magnetic resonance imaging (MRI) Dixon-based T1-weighted in-phase (IP)/opposed-phase (OP)/water-only (WO) imaging. Seventy-three patients with severe renal dysfunction (estimated glomerular filtration rate [eGFR] < 30 mL/min/1.73 m², CKD stage G4-5); 172 with moderate renal dysfunction (30 ≤ eGFR < 60 mL/min/1.73 m², CKD stage G3a/b); and 76 with mild renal dysfunction (eGFR ≥ 60 mL/min/1.73 m², CKD stage G1-2) participated in this study. The model was applied to the right, left, and both kidneys, as well as to each imaging method (T1-weighted IP/OP/WO images). The best performance was obtained when using bilateral kidneys and IP images, with an accuracy of 0.862 ± 0.036. The overall accuracy was better for the bilateral kidney models than for the unilateral kidney models. Our deep learning approach using kidney MRI can be applied to classify patients with CKD based on the severity of kidney disease.

Keywords: Chronic kidney disease; Deep learning; Dixon-based T1-weighted image; Magnetic resonance imaging; Three-dimensional convolutional neural network.

Figure 1

Flow chart of the inclusion and exclusion criteria for the study.

Figure 2

An overview of the image data processing used in this study. (A) On each Dixon-based T1-weighted kidney MRI, images are cropped to include the right and left kidneys; the resulting image volume has 24 coronal slices with each slice as 128 × 128 px. (B) These images are then further resized into 8 slices of 56 × 56 px for unilateral kidney datasets. Bilateral kidney data are obtained by stacking the data of the right and left kidneys, hence 16 slices of 56 × 56 px. Therefore, a total of 9 datasets are created for three-dimensional (3D) convolutional neural network (CNN) models derived (1) from each unilateral kidney (right or left kidney) and for bilateral kidneys, and (2) from each imaging method (T1-weighted in-phase (IP)/opposed-phase (OP)/water-only (WO) images), respectively. The 3D residual network-18 (3D ResNet-18)-based classification is performed on each dataset, classifying the three severity groups of chronic kidney disease (CKD).

Figure 3

The architecture of our three-dimensional residual network-18 model. Input is processed volumetric data of kidney magnetic resonance imaging. The network contains the initial convolutional layer, followed by 8 residual units (two with filters = 64, two with filters = 128, two with filters = 256, and two with filters = 512), each with two convolutional blocks as shown in the bottom row. The last layer is a fully connected dense layer that outputs a classification of three groups.

Figure 4

Confusion matrices show the status of multi-class classifications using bilateral (A–C), right (D–F) and left (G–I) kidney datasets with Dixon-based T1-weighted in-phase image (A, D, G), opposed-phase image (B, E, H), and water-only image (C, F, I), in classifying the three groups of chronic kidney disease. se-RD: severe renal dysfunction (estimated glomerular filtration rate [eGFR] < 30 mL/min/1.73 m²), mo-RD: moderate renal dysfunction (eGFR ≥ 30 and < 60 mL/min/1.73 m²), and mi-RD: mild renal dysfunction (eGFR ≥ 60 mL/min/1.73 m²). The data are means ± standard deviations.