Accurate identification of renal transplant rejection: convolutional neural networks and diffusion MRI

Abstract

For the past several years the ability of diffusion-weighted magnetic resonance imaging (DW-MRI) to provide a noninvasive assessment of renal transplant function has been investigated. The goal of this chapter is to develop a computer-aided diagnostic (CAD) system coupled with a deep convolutional neural network (DCNN) to help determine the functionality of renal transplant using diffusion MRI. This diffusion-MRI marker is derived from a 3D+ b-value DW-MRI. Our work includes kidney segmentation using a 3D DW-MRI, through a level-set approach aided by kidney/background appearance features as well as by shape. It also includes a feature extraction step in which the apparent diffusion coefficients (ADCs) of each voxel of the segmented DW-MRI at individual b-values are estimated, and lastly classification of renal transplant status. In addition, the utility of the extracted 3D ADCs for training and testing of the 3D DCNN–based classifier determines the status of the renal transplant. The results of the developed CAD system reached 94% accuracy, sensitivity, and specificity. We used the leave-one-out scenario as a cross-validation technique to determine acute-rejection versus nonrejection renal transplants. The conclusions ensure that the developed CAD system is highly reliable to diagnose the status of the renal transplant in a noninvasive way.