Renal segmentation from non-contrast T1-weighted MR images
Nicole Wake1, Jeremy C Lim2, Artem Mikheev1, Jas-mine Seah3, Elissa Botterill2, Shawna Farquharson4, Henry Rusinek1, and Ruth P Lim2,5

1Bernard and Irene Schwartz Center for Biomedical Imaging, Center for Advanced Imaging Innovation and Research, Department of Radiology, New York University School of Medicine, New York, NY, United States, 2Department of Radiology, Austin Health, Melbourne, Australia, 3Department of Endocrinology, Austin Health, Melbourne, Australia, 4Florey Neuroscience Institute, Melbourne, Australia, 5The University of Melbourne, Melbourne, Australia

Synopsis

A semi-automatic renal segmentation technique for non-contrast T1-weighted MR images was developed. Renal segmentation and volumetric analysis was tested in ten healthy volunteers and ten Type I diabetic patients. We found that this segmentation tool is fast, reliable, and requires minimal user interaction. Upon further validation, this method has clinical potential for monitoring renal status in appropriate patient populations.

Introduction

Kidney volume is an important measure for the early detection and monitoring of many renal diseases.1 Dynamic-contrast enhanced (DCE) imaging may be attractive for volumetric analysis since it provides exquisite contrast, however such DCE protocols emphasize temporal resolution at the cost of spatial resolution.2,3 Non-contrast T1-weighted gradient echo or 2D fast spin echo sequences are advantageous since they can provide high spatial resolution (1.5-2mm isotropic) within a single breath-hold. A fully-automated renal segmentation method effective for non-contrast MRI remains challenging due to respiratory motion artifacts and signal non-uniformity. To address this, we developed and validated a semi-automated renal segmentation approach.

Methods

Abdominal images obtained from ten healthy volunteers and ten Type I diabetic patients with early diabetic kidney disease were selected for volumetric analysis. All images were acquired on a 3T MRI (Skyra, Siemens, Erlangen, Germany) equipped with an 18-channel body phased-array coil anteriorly and a 32-channel spine coil posteriorly. Dixon volumetric interpolated breath-hold examination (VIBE) imaging was performed with a breath-hold at end-expiration: TR 3.8ms, TE 1.2ms (out of phase) and 2.5ms (in phase), FA 9°, FOV 384 × 384 × 208mm, true voxel size 2 × 2 × 3.8 mm3 interpolated to 2 × 2 × 2 mm3, acceleration factor 3 (GRAPPA), acquisition time 20s. Three observers with 13, 5, and 1 year of experience with renal MRI drew over-inclusive approximate contours around the kidney on 10% of slices. The effect of slice skip factor (s) was evaluated by adjusting the number of contours traced by each user from every four to fourteen slices. In order to complete the segmentation, a sequence of automated steps were performed by the segmentation program (http://wp.nyu.edu/firevoxel) including: (a) non-uniformity correction, (b) generating renal seed region, (c) selection of voxels with signal similar to seed signal (0.55 < S/S0 < 1.3), (d) hole filling and interpolation of contours and constrained 3D binary morphology.4 True renal masks were manually constructed by an abdominal radiologist experienced with renal anatomy and pathology. The accuracy of the segmentation tool was compared with reference masks, and with a recently developed tool, Robust Statistics Segmenter (RSS), based on the evolution of active contours.5

Results

There was a strong linear relationship between the estimated and true volumes calculated by the fully manual segmentation (adjusted R2 =0.97). Volume errors for the three readers were: 4.4 ± 3.0%, 2.9 ± 2.3%, and 3.1 ± 2.7%. The relative discrepancy between readers was 2.5 ± 2.1%. Figure 1 plots the volumes assessed by observers against the reference volume. Bland-Altman plots (Figure 2) showed a small (~4cm3) but systematic bias between R1 and the two other readers. The effect of slice skip factor (s) is shown in Figure 3. Volumetric errors increased by an average of 0.26% per unit increase in s between 4 and 10; and this error penalty increased to 1.57% in the range of 10-14. Segmentation using the optimized RSS algorithm was successful in 11/22 (55%) of cases with an error of 15.4± 6.1% and failed in 9/22 (45%) of cases. Failure was defined as a grossly inaccurate mask, with volume errors >25%. The average interactive processing time was 1.5 minutes per kidney. A representative resulting renal mask is shown in Figure 4.

Discussion and Conclusion

Segmentation from non-contrast MRI is challenging due to the similarities in gray-scale intensities of adjacent organs that may be in contact with the kidney. We have described a semi-automated segmentation tool that is fast, reliable and requires minimal user interaction. Upon further validation, this method has clinical potential for monitoring renal status in appropriate patient populations.

Acknowledgements

This work was supported by funding from the Diabetes Australia Research Trust and was performed under the rubric of the Center for Advanced Imaging Innovation and Research (www.cai2r.net), an NIBIB Biomedical Technology Resource Center (NIH P41 EB017183).

References

1. Zollner et al. Assessment of kidney volumes from MRI: acquisition and segmentation techniques. Am J Roentgenol. 1995; (5):1060-69.

2. Woodard et al. Segmental kidney volumes measured by dynamic contrast-enhanced magnetic resonance imaging and their association with CKD in older people. Am J Kidney Dis. 2015; 65(1):41-48.

3. Zollner et al. Assessment of 3D DCE-MRI of the kidneys using non-rigid image registration and segmentation. Comput Med Imaging Graph. 2012; 33(3):171-181.

4. Mikheev et al. Fully automatic segmentation of the brain from T1-weighted MRI using Bridge Burner algorithm. J Magn Reson Imaging. 2008; 27(6)1235-1241.

5. Gao et al. A 3D interactive multi-object segmentation tool using local robust statistics driven active contours. Med Image Anal. 2012; 16(6)1216-27.

Figures

Figure 1: Volume measured by three readers as compared to the true volume. Also shown are the regression line (solid) and the identity line (dashed).

Figure 2: Bland-Altman plots of the differences between volume estimations by pairs of readers. The solid lines show the mean difference between two readers (the bias), and the dotted lines indicate the 95% limits of agreement (± 2 standard deviations around the mean).

Figure 3: The effect of slice skip factor on segmentation error (red dots, left axis) and observer error (green dots, right axis).

Figure 4: Representative resulting renal mask. Red = semi-automated method; Yellow = reference mask; Orange = overlap of the two masks.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
4318