Estimation of single-kidney glomerular filtration rate (SK-GFR) using tracer kinetic model fitting to DCE-MRI data requires an arterial input function (AIF). Automated AIF voxel selection is required to produce a robust and practical tool for clinical use as manual or semi-automated selection is time consuming and can lead to inter-operator variability^{1, 2}. In this work we present an automatic AIF voxel selection method based on previous work³ that is tailored for renal DCE-MRI data. We assess its accuracy and precision against semi-automatic expert selection and reference nuclear medicine GFR values.A retrospective study was performed using data from 16 patients with reno-vascular disease. A 3D spoiled gradient echo sequence coronal acquisition (TR/TE/FA: 5.0ms/0.9ms/17°) was used to acquire a time-series with a temporal resolution of 2.1 s and a FOV / matrix size of 400x400x80 mm³ / 128x128x20. After AIF voxel selection was performed SK-GFR values were estimated by fitting the 2 compartment filtration model⁴ to average signal intensity change time curves for left and right whole kidney regions.A 3D rectangular region was drawn by the expert over the infrarenal segment (to minimise effects from inflow and turbulence near vessel junctions) on a map of maximum signal intensity change from average pre-contrast baseline (S_max – S0_mean). Software was then used to select voxels from the region within 5% of the maximum value.

The automated method was developed using distinct data sets from those presented in this work and was trained manually using visual inspection of the aorta seed, aorta mask, selected voxels and the AIF time-series.

1) A voxel in the aorta was detected from a map of maximum signal intensity change (maxSig: S_max – S0_mean) and a map of number of time points to peak signal intensity (TTP). A 3 by 11 kernel was used to generate maps of interquartile range (IQR) and mean for the maxSig and TTP maps. The IQR and mean maps were then normalised to values between 0 and 1 and the TTP map inverted. Scores were generated as follows: Score = (IQR_maxSig_norm + IQR_TTP_norm_inv)/2 * (mean_maxSig_norm + mean_TTP_norm_inv)/2. The voxel with the highest score was selected as the aorta seed (early high enhancing voxel surrounded by similar voxels in a long thin kernel).

2) A mask of the aorta was generated by flood filling from the aorta seed on the maxSig map until the region contained between 475 and 525 voxels.

3) The voxels in the mask were scored using the maxSig map to avoid partial voluming and a map of pre-contrast average signal intensity (aveBaseline: S0_mean) to avoid inflow effects (inflow appears as a higher pre-baseline signal). The maps were normalised to values between 0 and 1 and the aveBaseline map inverted. Scores were generated for each voxel as follows: maxSig_norm * aveBaseline_inv_norm. The voxels within a percentage threshold of the maximum score were selected. Three percentage thresholds were applied: 5%, 10% and 25%.

From visual inspection, all voxels selected by the automated method were seen to lie within the aorta. Figure 1 compares the automated methods to the expert method. The 25% threshold has the lowest bias (-0.5 ml/min) and limits of agreement (-7.13 to 6.11 ml/min). Figure 2 compares the expert and automated methods to nuclear medicine values. The bias and limits of agreement for the 25% threshold are within 2 ml/min of the expert method. Figure 3 shows the selected voxels and time courses for the greatest outlier when comparing the automated to the expert method. This is also representative of the second greatest outlier. Figure 4 shows the third greatest outlier.The expert and automated method with a 25% threshold show similar accuracy and precision when compared to the nuclear medicine values. However, poor agreement between the expert and automated methods is sometimes seen. The two greatest outliers (Figure 3) have a few voxels with very high peaks near the renal arteries which are selected by the automated but not the expert method. For the third greatest outlier (Figure 4), the expert AIF and higher automated thresholds may be affected more by partial voluming (lower peaks). However, a partial volumed AIF would not be expected to give a closer match to the reference values. This may be due to other issues such as motion.The automated voxel selection method shows promise as a robust tool for clinical use as it consistently selects voxels within the aorta and is easy to use. However, further investigation is required to understand the effect of confounding factors such as partial voluming and motion.