Premium
Automatic intra‐subject registration‐based segmentation of abdominal fat from water–fat MRI
Author(s) -
Joshi Anand A.,
Hu Houchun H.,
Leahy Richard M.,
Goran Michael I.,
Nayak Krishna S.
Publication year - 2013
Publication title -
journal of magnetic resonance imaging
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.563
H-Index - 160
eISSN - 1522-2586
pISSN - 1053-1807
DOI - 10.1002/jmri.23813
Subject(s) - abdominal fat , segmentation , subject (documents) , radiology , medicine , computer science , nuclear medicine , artificial intelligence , obesity , library science
Abstract Purpose: To develop an automatic registration‐based segmentation algorithm for measuring abdominal adipose tissue depot volumes and organ fat fraction content from three‐dimensional (3D) water–fat MRI data, and to evaluate its performance against manual segmentation. Materials and Methods: Data were obtained from 11 subjects at two time points with intermediate repositioning, and from four subjects before and after a meal with repositioning. Imaging was performed on a 3 Tesla MRI, using the IDEAL chemical‐shift water–fat pulse sequence. Adipose tissue (subcutaneous—SAT, visceral—VAT) and organs (liver, pancreas) were manually segmented twice for each scan by a single trained observer. Automated segmentations of each subject's second scan were generated using a nonrigid volume registration algorithm for water–fat MRI images that used a b‐spline basis for deformation and minimized image dissimilarity after the deformation. Manual and automated segmentations were compared using Dice coefficients and linear regression of SAT and VAT volumes, organ volumes, and hepatic and pancreatic fat fractions (HFF, PFF). Results: Manual segmentations from the 11 repositioned subjects exhibited strong repeatability and set performance benchmarks. The average Dice coefficients were 0.9747 (SAT), 0.9424 (VAT), 0.9404 (liver), and 0.8205 (pancreas); the linear correlation coefficients were 0.9994 (SAT volume), 0.9974 (VAT volume), 0.9885 (liver volume), 0.9782 (pancreas volume), 0.9996 (HFF), and 0.9660 (PFF). When comparing manual and automated segmentations, the average Dice coefficients were 0.9043 (SAT volume), 0.8235 (VAT), 0.8942 (liver), and 0.7168 (pancreas); the linear correlation coefficients were 0.9493 (SAT volume), 0.9982 (VAT volume), 0.9326 (liver volume), 0.8876 (pancreas volume), 0.9972 (HFF), and 0.8617 (PFF). In the four pre‐ and post‐prandial subjects, the Dice coefficients were 0.9024 (SAT), 0.7781 (VAT), 0.8799 (liver), and 0.5179 (pancreas); the linear correlation coefficients were 0.9889, 0.9902 (SAT, and VAT volume), 0.9523 (liver volume), 0.8760 (pancreas volume), 0.9991 (HFF), and 0.6338 (PFF). Conclusion: Automated intra‐subject registration‐based segmentation is potentially suitable for the quantification of abdominal and organ fat and achieves comparable quantitative endpoints with respect to manual segmentation. J. Magn. Reson. Imaging 2013;37:423–430. © 2012 Wiley Periodicals, Inc.