Glomerular filtration rate (GFR) can be estimated using dynamic contrast enhanced MRI (DCE-MRI) and pharmacokinetic models^1-3. The pharmacokinetic models calculate the GFR from the signal intensity changes in the aorta and the renal cortex, obtained from DCE-MRI data by segmenting these regions of interest. Manual segmentation of kidneys can take several hours, adding cost, and many clinics lack personnel for this labor-intensive task. Here, we describe a novel method for automatic renal segmentation based on morphological segmentation and machine learning, and assess the performance of the method.

Automated segmentation approach: The segmentation of the renal cortex is performed in two steps (Fig.1). The first step is the segmentation of the kidney from the abdominal images using morphological segmentation methods. This process starts by creating a volumetric image dataset by applying maximum intensity projection in time (MIP-t). As kidneys typically enhance intensely, MIP-t creates high contrast between kidneys and the surrounding tissues (Fig.2a). Then, a local intensity dependent binary thresholding method (Local Otsu) is applied to all slices in axial, sagittal and coronal planes removing the dark background and leaving only the bright regions (Fig.2b). The connected components are identified in the 3D volume and treated as an object. Small connections between large objects are broken using watershed segmentation in coronal projections (Fig.2c). The result of the morphological segmentation step is a series of 3D objects that correspond to the high intensity regions of the MR images.

The second step is the segmentation of the kidney into cortex and medullary regions using machine learning techniques. First, a support vector classifier (SVC) is used to identify the kidney parenchymal voxels in the abdominal images. Classification was performed on each voxel using 6 features: 3 features for position (x, y, z) and 3 features for percent signal change. Percent signal change ($\Delta{S}$) features are calculated using $\Delta{S}(t)=\frac{S(t)-S(0)}{S(0)}$ where $S(t)$ is the signal intensity of the voxel at time $t$. Dimensionality reduction (PCA) is applied on $\Delta{S}$ to reduce it to 3 dimensions. The result of this classification is used for identification of kidney objects in the morphological segmentation output. Once the kidney objects are identified, another SVC is applied to these objects to separate cortex from medulla. This classification is performed on each voxel within the kidney object using 53 features: 3 for position, 25 for percent signal change and 25 for signal intensity over time with no dimensionality reduction.

11 consecutive pediatric subjects clinically referred for contrast enhanced MRI on a GE 3T scanner were retrospectively identified with IRB approval. Acquisitions were performed using VDRad⁴ with parameters: 15° flip angle, ±100kHz bandwidth, TR = 3.3ms, matrix = 192x180, FOV = 320x256 mm, slice thickness = 2.4 mm, 80 slices, 36 temporal phases, 3s temporal resolution and 6.2x acceleration, 32 channel torso coil. Injection protocol: single dose contrast diluted to 10ml was power injected 
at a rate of 1ml/s. Manual segmentations of images were performed before training our automatic segmentation models. Training and testing was performed using a leave one out cross-validation (LOOCV). Segmentation results were used to calculate the transfer coefficients (Ktrans) of kidneys using a 2-compartment model². The pharmacokinetic analysis was performed using a common aortic input function between automatic and manual renal segmentations. Automatic segmentation results were compared to the manual segmentation maps using precision (PPV) and recall metrics (sensitivity) as well as Ktrans values.Some of the segmentation maps generated by this method are shown in Fig.3. The results show that the method in general provides good renal cortex coverage (recall=0.92±0.06) but misclassifies some medulla voxels as cortex (precision=0.83±0.08). These misclassifications cause very little to no change in the average signal intensity of the renal cortex. The average estimated Ktrans was 0.23±0.6min^-1 using the automated method and 0.23±0.7min^-1 using the manual method. However, the low precision causes large errors in estimated renal cortex volume which is needed for GFR estimation. Total renal volume is more accurate because it isn’t affected by the cortex-medulla misclassifications. Hence, we trained a linear regression model using RANSAC to predict the cortex volume from total renal volume yielding a 44% reduction in GFR RMSE (Fig.4). The mean error in the cortical volume estimates after the correction, 6.2±5.9ml, was below 16.3±11.2ml the observed average error between two human experts⁵. The GFR estimates and split function were computed after volume correction (Fig.5). We have presented a fully automatic renal segmentation method for detecting renal cortex and showed that renal filtration parameter estimates are very similar to the manual segmentation.