Evaluation of Two Deformable Registration Algorithms for Assessment of Brown Adipose Tissue in Humans
Vanessa Stahl1, Martin T. Freitag2, Armin M. Nagel1,3, Ralf O. Floca4, Moritz C. Berger 1, Jan P. Karch5, Peter Bachert1, Mark E. Ladd1, and Florian Maier1

1Medical Physics in Radiology, German Cancer Research Center, Heidelberg, Germany, 2Department of Radiology, German Cancer Research Center, Heidelberg, Germany, 3Department of Diagnostic and Interventional Radiology, University Medical Center Ulm, Ulm, Germany, 4Medical and Biological Informatics, German Cancer Research Center, Heidelberg, Germany, 5Institue of Physics, Johannes Gutenberg University Mainz, Mainz, Germany


Human brown adipose tissue (BAT) is mostly found in cervical and mediastinal anatomic sites, making MR-imaging challenging because of susceptibility to breathing motion artifacts. Image acquisition under breath-hold requires data registration, especially for long measurement times. Two deformable registration algorithms (Fast Symmetric Forces Demons (FSF), Level Set Motion (LSM)) were evaluated regarding their suitability for compensation of deviations in breath-hold positions. Data processing was based on a volunteer study using distinct anatomical landmarks placed by an experienced radiologist. Landmark positions were evaluated after transformation, showing that FSF is more suitable for registration of thoracic data allowing for human BAT assessment.


One aspect of ongoing research on obesity and metabolic diseases1 is the assessment of human brown adipose tissue (BAT) using MRI2-4. Imaging is challenging because human BAT is a heterogeneous mixture5 of brown and white adipocytes. Isotropic image resolutions below (1.5 mm)3 are required to distinguish between both. Additionally, BAT has been mostly observed in the cervical and mediastinal anatomic areas6,7, which results in a high susceptibility to breathing motion during image acquisition. One option is to acquire images during breath-hold. However, breath-hold positions differ due to long measurement times requiring data registration. In this study, two deformable registration algorithms were evaluated with respect to their suitability for compensation of deviations in breath-hold positions during BAT assessment.

Material and Methods

Data from a previously published human BAT assessment study8 were analyzed retrospectively. The study was performed using a 3T MRI system (Biograph mMR 3.0T, Siemens, Erlangen, Germany) with five healthy volunteers. 2-point Dixon9 imaging (VIBE; TR=5.85ms, TEin/TEopp=2.46/3.69ms, α=10° matrix=416×260×64, FOV=500×312×76.8mm3, BW=710 Hz/px, GRAPPA R=2; TA=30sec) was used during breath-hold. Water and fat images were acquired with a 5-minute time resolution for 140 minutes. Registration of data sets to a target data set (DT, water image) was done with MITK (Medical Imaging Interaction Toolkit, DKFZ, Heidelberg, Germany)10. The Insight Tookit (itk)-based11 implementations of the deformable Fast Symmetric Forces Demons12 (FSF) and Level Set Motion13 (LSM) algorithms were used. For each volunteer, a data set at the beginning (DM1) and at the end of the measurement (DM2) were chosen. Water images of DM1 and DM2 were registered to the target data DT using FSF and LSM resulting in transformations TFSF and TLSM, respectively (Fig.1a). Registration evaluation was done using PointSet Interaction (plugin of MITK). A radiologist (5 years of experience) placed a set of five positions XT={xi ǀ i=1…5} at distinct anatomical landmarks on DT. The following landmarks were chosen (Fig.1b): L1: joint space of the right clavicle (axial), L2: branching of the left common carotid artery from the aortic arch (coronal), L3: left costo-vertebral joint space of the 1st thoracic vertebra (axial), L4: processus spinosus of the 7th cervical vertebra (sagittal), L5: left tuberculum majus humeri (sagittal). The positions XM1={xi,M1 ǀ i=1…5} and XM2={xi,M2 ǀ i=1…5} of the same landmarks were marked on the unregistered data sets (DM1, DM2). Using TFSF and TLSM, both sets of positions XM1 and XM2 were mapped to the target. Deviation vectors ∆si = xi,M1/2 - xi,T for all landmarks Li (i=1…5) for all volunteers were calculated. Mean deviations ∆si,mean for each landmark were compared for both algorithms.


Boxplots show the distribution range of each mapped landmark for FSF and LSM (Fig.2). Median values scatter around zero in the range of MFSF=[-0.70; 0.31]mm and MLSM=[-2.91; 1.22]mm. For each landmark, a shift of a single voxel after (side length 1.2mm) after mapping with FSF and LSM is illustrated (Fig.3). Mean deviations ∆si,mean scatter around zero in the range [-0.90; 0.51]mm for FSF and [-2.49; 1.06]mm for LSM. LSM results show larger maximum deviations and larger absolute deviations compared to FSF. In general, landmark L5 shows the maximum absolute deviation, with ∆s5,FSF=1.61mm≙1.34pixel and ∆s5,LSM=2.64mm≙-2.20pixel. Combining the results for all landmarks L1-L5 and focusing on the directions in space (x-,y-,z-components), mean deviation vectors are ∆mFSF=(-0.18; -0.19; -0.26)mm and ∆mLSM=(-0.21; -0.59; -0.51)mm (Fig.4,table).

Discussion and Conclusion

Neither registration algorithm (FSF and LSM) showed a systematic displacement of the landmarks. Both deviation sets ∆si,mean scatter around zero. FSF results show smaller mean deviations and smaller maximum absolute deviations. Hence, with the limited number of landmarks, FSF seems to be suitable for correction of BAT data. Considering the evaluated mean BAT volume of 0.42ml in a recent volunteer study8 and assuming a simplified cubical shape (7.4mm≙6.2pixel side length), a shift using the mean deviations ∆mFSF and ∆mLSM still results in an overlap of OVmean,FSF=92% and OVmean,LSM=83% (Fig.4a). Considering the maximum deviations of directions in space, no overlap (OVall) exists (Fig.4b). High deviations of landmark L5 lead to the exclusion of BAT assessment in this site. Hence, the maximum error including L1-L4 leads to an overlap of OVexL5=25% for FSF mapping (Fig.4c). L1, L3, L4 are close to anatomical BAT sites1,6,7. Although the volunteer study8 focused on interscapular BAT, the results for FSF allow for evaluation of more BAT sites, e.g. supraclavicular BAT. Since supraclavicular BAT depot volumes are expected to be larger, an increase in overlap is expected. In conclusion, this study indicates that FSF is suitable for correction of breath-hold offsets in human data and enables improved assessment of human BAT depots.


This work was funded by the Helmholtz Alliance ICEMED - Imaging and Curing Environmental Metabolic Diseases, through the Initiative and Networking Fund of the Helmholtz Association.


1. Berriel Diaz et al, Metabolism 63: 1238-49 (2014)

2. Chen et al., JNM, 54: 1584-1587 (2013)

3. Rasmussen et al., PLoS ONE 8: e77907 (2013)

4. Hu et al., Magn Reson Imaging 32: 107-117 (2014)

5. Lidell et al., Nature Medicine 19: 631-634 (2013)

6. Nedergaard et al., Am J Physiol Endocrinol Metab 293: E444-E452 (2007)

7. Cypess et al., N Engl J Med 360: 1509-17 (2009)

8. Stahl et al., ISMRM 2015, 336

9. Dixon et al., Radiology 153: 189-194 (1984)

10. Nolden et al., Int J Comput Assist Radiol Surg, 8: 607-20 (2013)

11. Vercauteren et al., Insight Journal, ISC/NA-MIC Workshop on Open Science at MICCAI (2007)

12. Vercauteren et al., NeuroImage 45: S61–S72 (2009)

13. Vemuri et al., Medical Image Analysis 7: 1-20 (2003)


(a) Registration of unregistered data sets (DM1, DM2) to target data (DT) using FSF and LSM. Applications of algorithms result in transformations TFSF and TLSM. (b) Overview of anatomical sites for landmarks L1-L5 set by radiologist for evaluation. Water signal outside of body due to water tubes of cooling circuit8.

Boxplot illustration for each component of deviation sets ∆si for FSF and LSM. Median values for FSF and LSM both scatter around zero. Range of MFSF is narrower, and results exhibit smaller deviations (smaller whiskers) for FSF compared to LSM. Median values are given in table.

Illustration of resulting deviations ∆si with the help of a single voxel (1.2×1.2×1.2mm³) in the target data DT (blue) considered for each landmark L1-L5. Mapping the positions XM1/M2 on the target using FSF results in smaller deviations (green) compared to LSM mapping (red). Mean deviations±standard deviation given in table.

Mean evaluated BAT volume (0.42ml, 7.4×7.4×7.4mm³) illustrated as cubical shape. Volumes overlap after shifts by mean deviations (a,OVmean), maximum deviations (b,OVall) and maximum deviations excluding L5 (c,OVexL5). Mean deviations±standard deviation and overlap ratios given in table.

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