A Novel CNN-Based CAD System for Early Assessment of Transplanted Kidney Dysfunction

Abstract

This paper introduces a deep-learning based computer-aided diagnostic (CAD) system for the early detection of acute renal transplant rejection. For noninvasive detection of kidney rejection at an early stage, the proposed CAD system is based on the fusion of both imaging markers and clinical biomarkers. The former are derived from diffusion-weighted magnetic resonance imaging (DW-MRI) by estimating the apparent diffusion coefficients (ADC) representing the perfusion of the blood and the diffusion of the water inside the transplanted kidney. The clinical biomarkers, namely: creatinine clearance (CrCl) and serum plasma creatinine (SPCr), are integrated into the proposed CAD system as kidney functionality indexes to enhance its diagnostic performance. The ADC maps are estimated for a user-defined region of interest (ROI) that encompasses the whole kidney. The estimated ADCs are fused with the clinical biomarkers and the fused data is then used as an input to train and test a convolutional neural network (CNN) based classifier. The CAD system is tested on DW-MRI scans collected from 56 subjects from geographically diverse populations and different scanner types/image collection protocols. The overall accuracy of the proposed system is 92.9% with 93.3% sensitivity and 92.3% specificity in distinguishing non-rejected kidney transplants from rejected ones. These results demonstrate the potential of the proposed system for a reliable non-invasive diagnosis of renal transplant status for any DW-MRI scans, regardless of the geographical differences and/or imaging protocol.

Figure 1

The proposed convolutional neural network (CNN)-based framework for early detection of renal transplant rejection using diffusion-weighted (DW) MRI. This system consists of three main processing steps. In the first step, the histogram of input DW-MRI data is equalized to reduce the noise and inhomogeneity of intensities. Then, an ROI enclosing the kidney in each subject is constructed. In the second step, the 3D ADC maps are estimated for the selected ROI and then fused with the clinical biomarkers, i.e., the creatinine clearance (CrCl) and serum plasma creatinine (SPCr), for allografts classification. In the final step, the fused markers are fed as a 3D input of size 150 × 150 × 24 voxels to the proposed CNN-based classifier to classify renal allografts into non-rejection (NR) and acute rejection (AR).

Figure 2

Illustration of the voxel-wise ADC calculations at a voxel (x; y; z) at b-value of 500 s mm⁻²: (a) the ROI-kidney regions at b₀, (b) the ROI-kidney regions at b₅₀₀, and (c) the constructed ADC maps for the defined ROI.

Figure 3

Demonstration of the efficacy of the fusion of both image and clinical markers. In this work, the estimated ADC maps are fused with the one dimensional (1D) CrCl and 1D SPCr values obtained during routine post-transplantation monitoring. The clinical biomarkers of each subject were first normalized with regard to the maximum values of each marker. Then, the normalized values were added to the voxel-wise ADC maps at all of the b-value scans. This difficult example differentiates an NR case (a) from an AR case (b). As illustrated in this figure, it is very difficult using the ADC maps alone to distinguish between the normal and abnormal subjects. This can be justified by the large overlap of the ADC values between these two subjects, which have been revealed by the color-coded maps. Visually, it is clear that the two subjects had a good color separation after the fusion process of both markers, where the dark-green color represents poor kidney function (i.e. low ADCs + low CrCl + high SPCr) and the orange-yellowish color represents a normal kidney function (i.e. high ADCs + high CrCl + low SPCr).

Figure 4

Illustration of the processing of a single volume using CNN. To clarify how a single convolution layer of the CNN processes a 3D input volume of size 150 × 150 × 24 voxels, each 3D volume has 24 images and a 2D convolution is applied to each image using a kernel of size K × K. The output of an individual kernel is then summed to create a 2D feature map. Usually, multiple kernels are used, and the above step is applied again for those kernels, and different feature maps are produced for respective kernels. The final result is a volume of feature maps.

Figure 5

The architecture of the proposed convolutional neural network (CNN), where (a) illustrates the configurations of the developed CNN and its layers and (b) represents the pipeline of the fusion process using a support vector machine (SVM) classifier. The CNN is trained and validated at each b-value apart from the other b-values. At each b-value, we have 56 samples. Each sample is a 3D volume for a certain subject, and each volume is 150 × 150 × 24 voxels. Each 3D volume is fed to a CNN as one sample of the dataset. The final classification decision is obtained by combining the decisions of all CNNs’ output probabilities at all b-values. This fusion is achieved by a support vector machine (SVM) where each sample has 22 features (2 probabilities for the two classes obtained from the CNN at a specific b-value × 11 b-values).

Figure 6

Receiver operating characteristics (ROC) curves for the FBio scenario for individual b-values and their fusion.

Figure 7

Receiver operating characteristics (ROC) curves for the proposed CNN-based system (for both scenarios S₁ and S₂), the clinical biomarkers (ClinBio) based upon using a support vector machine (SVM) classifier, and the auto-encoding system.