Automated Segmentation of Kidneys from MR Images in Patients with Autosomal Dominant Polycystic Kidney Disease

Abstract

Background and objectives Our study developed a fully automated method for segmentation and volumetric measurements of kidneys from magnetic resonance images in patients with autosomal dominant polycystic kidney disease and assessed the performance of the automated method with the reference manual segmentation method.Design, setting, participants, & measurements Study patients were selected from the Consortium for Radiologic Imaging Studies of Polycystic Kidney Disease. At the enrollment of the Consortium for Radiologic Imaging Studies of Polycystic Kidney Disease Study in 2000, patients with autosomal dominant polycystic kidney disease were between 15 and 46 years of age with relatively preserved GFRs. Our fully automated segmentation method was on the basis of a spatial prior probability map of the location of kidneys in abdominal magnetic resonance images and regional mapping with total variation regularization and propagated shape constraints that were formulated into a level set framework. T2-weighted magnetic resonance image sets of 120 kidneys were selected from 60 patients with autosomal dominant polycystic kidney disease and divided into the training and test datasets. The performance of the automated method in reference to the manual method was assessed by means of two metrics: Dice similarity coefficient and intraclass correlation coefficient of segmented kidney volume. The training and test sets were swapped for crossvalidation and reanalyzed.Results Successful segmentation of kidneys was performed with the automated method in all test patients. The segmented kidney volumes ranged from 177.2 to 2634 ml (mean, 885.46569.7 ml). The mean Dice similarity coefficient 6SD between the automated and manual methods was 0.8860.08. The mean correlation coefficient between the two segmentation methods for the segmented volume measurements was 0.97 (P,0.001 for each crossvalidation set). The results from the crossvalidation sets were highly comparable.Conclusions We have developed a fully automated method for segmentation of kidneys from abdominal magnetic resonance images in patients with autosomal dominant polycystic kidney disease with varying kidney volumes. The performance of the automated method was in good agreement with that of manual method.

Comparison of Total Kidney Volume Quantification Methods in Autosomal Dominant Polycystic Disease for a Comprehensive Disease Assessment

“…The results of this comparison are in line with the ones previously presented [3,4,5,6,8,13,14]. Slight differences were found for variability analysis that could be explained by considering that the population selected for this analysis is characterized by huge kidneys.…”

“…It was not possible to perform a direct quantitative comparison with the method presented in [8], since a specific algorithm was not available. The method in [8] was tested on a population of smaller kidneys (TKV range: 177-2,634, mean TKV: 885 ± 570) and validated versus manual tracing.…”

“…The method in [8] was tested on a population of smaller kidneys (TKV range: 177-2,634, mean TKV: 885 ± 570) and validated versus manual tracing. Compared to the highly automated method, linear regression coefficient and ICC were slightly minor ( r = 0.97, ICC = 0.97).…”

“…These matters justify several recent studies aimed at the computation of TKV in ADPKD patients from MRI [3,4,5,6,7,8]. A comprehensive review of the state-of-the-art MRI for the estimation of kidney volume in chronic kidney disease is presented in [9].…”

“…These methods have differences between them, since some rely on geometric approximations [3,6] limiting manual contouring to one mid-MRI slice; others are highly automated [4,5] requiring a single seed point selection in one mid-MR slice to obtain the 3D kidney surfaces. Very recently, two additional fully automated approaches were proposed; the first one [7] was exclusively proposed to monitor disease progression starting from patient kidney surfaces obtained by manual contouring during the first visit; the second one [8] was based on spatial probability density maps and regional mapping with total variation regularization and propagated shape constraints.…”