Deep Convolutional Neural Networks in Application to Kidney Segmentation in the DCE-MR Images

This paper evaluates three convolutional neural network architectures – U-Net, SegNet, and Fully Convolutional (FC) DenseNets – in application to kidney segmentation in the dynamic contrast-enhanced magnetic resonance images (DCE-MRI). We found U-Net to outperform the alternative solutions with the Jaccard coefficient equal to 94% against 93% and 91% for SegNet and FCDenseNets, respectively. As a next step, we propose to classify renal mask voxels into cortex, medulla, and pelvis based on temporal characteristics of signal intensity time courses. We evaluate our computational framework on a set of 20 DCE-MRI series by calculating image-derived glomerular filtration rates (GFR) – an indicator of renal tissue state. Then we compare our calculated GFR with the available ground-truth values measured in the iohexol clearance tests. The mean bias between the two measurements amounts to −7.4 ml/min/1.73 m2 which proves the reliability of the designed segmentation pipeline.