European Radiology

, Volume 23, Issue 1, pp 246–252 | Cite as

Image registration improves human knee cartilage T1 mapping with delayed gadolinium-enhanced MRI of cartilage (dGEMRIC)

  • Esther E. Bron
  • Jasper van Tiel
  • Henk Smit
  • Dirk H. J. Poot
  • Wiro J. Niessen
  • Gabriel P. Krestin
  • Harrie Weinans
  • Edwin H. G. Oei
  • Gyula Kotek
  • Stefan Klein
Open Access
Musculoskeletal

Abstract

Objectives

To evaluate the effect of automated registration in delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) of the knee on the occurrence of movement artefacts on the T1 map and the reproducibility of region-of-interest (ROI)-based measurements.

Methods

Eleven patients with early-stage knee osteoarthritis and ten healthy controls underwent dGEMRIC twice at 3 T. Controls underwent unenhanced imaging. ROIs were manually drawn on the femoral and tibial cartilage. T1 calculation was performed with and without registration of the T1-weighted images. Automated three-dimensional rigid registration was performed on the femur and tibia cartilage separately. Registration quality was evaluated using the square root Cramér–Rao lower bound (CRLBσ). Additionally, the reproducibility of dGEMRIC was assessed by comparing automated registration with manual slice-matching.

Results

Automated registration of the T1-weighted images improved the T1 maps as the 90% percentile of the CRLBσ was significantly (P < 0.05) reduced with a median reduction of 55.8 ms (patients) and 112.9 ms (controls). Manual matching and automated registration of the re-imaged T1 map gave comparable intraclass correlation coefficients of respectively 0.89/0.90 (patients) and 0.85/0.85 (controls).

Conclusions

Registration in dGEMRIC reduces movement artefacts on T1 maps and provides a good alternative to manual slice-matching in longitudinal studies.

Key Points

Quantitative MRI is increasingly used for biomedical assessment of knee articular cartilage

Image registration leads to more accurate quantification of cartilage quality and damage

Movement artefacts in delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) are reduced

Automated image registration successfully aligns baseline and follow-up dGEMRIC examinations

Reproducibility of dGEMRIC with registration is similar to that using manual slice-matching

Keywords

Magnetic resonance imaging Articular cartilage Knee osteoarthritis Computer-assisted image processing Reproducibility of results 

Abbreviations and acronyms

CRLB

Cramér–Rao lower bound

CRLBσ

Square root of the Cramér–Rao lower bound

90%-CRLBσ

90% percentile of the CRLBσ in a region of interest

CV

Coefficient of variation

dGEMRIC

Delayed gadolinium-enhanced MRI of cartilage

FSPGR

Fast spoiled gradient-recalled echo

VGA

Visual grading analysis

Introduction

Osteoarthritis, the most common form of arthritis, is characterised by the degradation and loss of cartilage [1]. Delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) has been introduced as a non-invasive quantitative technique to measure cartilage quality by evaluating its glycosaminoglycan content [2, 3, 4]. Imaging is performed after administration of a negatively charged contrast agent that diffuses into cartilage in reverse relationship to the charge of glycosaminoglycan molecules. A quantitative T1 map is reconstructed from images acquired consecutively with different inversion times (TIs).

Automated image registration may improve dGEMRIC studies in two ways. First, T1 map quality may be improved by registration of the T1-weighted images to correct for patient movement during the acquisition process. Only a few papers on dGEMRIC used image registration to correct for motion artefacts [5, 6]. Moreover, registration of the T1-weighted images has not been evaluated before in patients with knee osteoarthritis. Second, automated image registration can be applied in longitudinal studies to compare T1 maps acquired at different time points. Previous studies did not use registration for this purpose, but instead relied on manual matching of slices [7, 8], which is operator-dependent and time-consuming. Automated image registration eliminates the subjective manual slice-matching and manually outlined ROIs are required only once. This can be of value in follow-up studies evaluating for example the effect of a treatment.

The present study evaluates the effects of both image registration steps on the outcome of dGEMRIC in patients with knee osteoarthritis as well as in healthy controls.

Materials and methods

Subjects

This study was approved by the institutional review board of Erasmus MC, Rotterdam, The Netherlands. Informed consent was obtained from all subjects.

Two subject groups were included in this study. Group I consisted of 11 patients (age 52.2 ± 11.0 years, seven male) with early-stage knee osteoarthritis (knee complaints > 3 months, visual analogue pain scale > 20 mm and Kellgren–Lawrence grade I or II on radiography [9]). Group II consisted of ten healthy controls (age 26.7 ± 8.6 years, four male).

Imaging protocol

Images were acquired on a 3-T MRI (Discovery MR750; General Electric Healthcare, Milwaukee, WI, USA). The three-dimensional (3D) protocol for T1-weighted imaging consisted of an inversion recovery fast spoiled gradient-recalled echo (FSPGR) sequence with five TIs (100; 200; 400; 800; 2,100 ms) [10]. The repetition time (TR) was the inversion time plus the time after read-out (TS = 3.9 ms). Other parameters were: flip angle =15°, echo time = 1.5 ms, field of view = 15 × 15 cm, slice thickness = 3 mm, slice spacing = 3 mm, in-plane voxel size = 0.6 × 0.6 mm, number of slices in the sagittal plane = 36. The dGEMRIC MR protocol lasted approximately 15 min.

In patients, the dGEMRIC protocol [10] was performed twice with an interval of 7 days (range 5–14 days). A double dose (0.2 mmol/kg) of Magnevist (Bayer, Berlin, Germany) was injected intravenously. Next, the patients were asked to cycle for 10 min on a home trainer to promote contrast agent distribution into and throughout the knee and the cartilage [11]. After cycling, there was a delay of 80 min before the participants underwent MRI. An open design three-channel knee coil (Flick Engineering Solutions, Winterswijk, The Netherlands) was used, which enabled imaging of patients with a large knee diameter. Controls underwent MRI with the standard eight-channel knee coil (General Electric Healthcare) requiring a knee diameter less than 14 cm. For the controls, no contrast agent was used, and the second MR examination was acquired after a short break and repositioning of the knee.

Definition of regions of interest

For each subject, two regions of interest (ROIs) on the femoral and tibial cartilage were outlined by a trained researcher with a medical degree (J.vT.). The femoral cartilage ROI consisted of the adjacent trochlear, weight-bearing and posterior cartilage of the femur, and the tibial cartilage ROI consisted of the weight-bearing tibial plateau cartilage. The ROI outlining was performed on the central slice through the medial and lateral tibiofemoral joint, for both the first and second MR examination, resulting in eight cartilage ROIs per subject.

T1 calculation and uncertainty estimate

The T1 map was reconstructed by voxelwise fitting of the relationship SI(TI) = S0·(1-A·exp(-TI/T1) + exp(-TR/T1)) [10] to the T1-weighted images acquired at a range of inversion times. The fitting was performed with a maximum likelihood estimator of T1, S0 (fully relaxed signal) and A (inversion efficiency), which takes into account the Rician distribution of the data because, for magnitude MR images, this is more accurate than the commonly used normal distribution [12].

The uncertainty of the estimated parameters at each voxel can be expressed by the Cramér–Rao lower bound (CRLB), which gives a lower bound for the variance [12, 13, 14]. The square root of the T1 CRLB (CRLBσ) can therefore be interpreted as a lower bound for the standard deviation of the T1 value, which quantifies how noise on the MR signal propagates to uncertainty of the estimated T1 value. In quantitative MRI, CRLBσ has previously been used for optimisation of MR sequences [15, 16, 17], but it can also be used as an indicator of misalignment. Misalignment of the T1-weighted images, especially at tissue boundaries, results in biologically implausible values of S0, A and T1, often associated with a high uncertainty, which is expressed by CRLBσ. The T1 calculations were performed using in-house developed Matlab software (R2008a; The MathWorks, Natick, MA, USA), which produces both the T1 map and the CRLBσ map. As a summary statistic for the CRLBσ values, we computed the 90% percentile (90%-CRLBσ) over all voxels in each annotated ROI; the lower this value, the better. We also computed the 90%-CRLBσ over all voxels in all ROIs together, to obtain a single measure per subject.

Registration of T1-weighted images

All T1-weighted images were registered in 3D with respect to that T1-weighted image showing the highest contrast between cartilage and surrounding synovial fluid, and between cartilage and bone cortex (FSPGRTI = 2,100). Registrations were performed using Elastix software [18] using a rigid transformation model (translations and rotations). Femoral and tibial regions were registered separately based on subvolumes containing only the specific bone and surrounding tissue to allow correction for motion of the knee joint (Fig. 1). The registration was optimised over 1,000 iterations with localised mutual information (LMI) as a similarity measure [19]. Per iteration, LMI was calculated using 2,048 random samples obtained from a sample region of size 50 × 50 × 50 mm. Cubic B-spline interpolation was used when applying the deformation to the moving image. The exact registration settings can be found on the parameter file database on the Elastix website: http://elastix.isi.uu.nl/wiki.php.
Fig. 1

Registration is performed separately on the femoral (red) and tibial (blue) subvolumes to correct for motion of the knee joint. The background greyscale image is a T1-weighted image (FSPGRTI = 2,100)

T1 maps were calculated with and without registration. A Wilcoxon signed rank test was used to test for a significant effect of registration on the 90%-CRLBσ values.

Registration between the first and second MR examinations

To align the FSPGRTI = 2,100 images from the first and second MR examinations, registration was performed with the same method as in the previous section. Based on this registration, the T1 maps from the second MR examination were transformed to the T1 maps obtained in the initial study. The result of this alignment was compared with the reference standard of manually selecting matching slices, which was performed by a trained researcher (J.vT.) visually inspecting FSPGRTI=2,100 images from the first and second MR examinations for matching slices.

Analysis of the reproducibility between the first and second MR examinations was based on correlations of a weighted mean T1 value per ROI. A weighted mean was computed to reduce the effect of outliers which are, for example, caused by bone voxels accidently included in the cartilage ROIs. As outliers are expected to have a high CRLBσ, the reciprocal of the CRLBσ was used as the weight of each voxel to reduce their effect.

The reproducibility of the weighted mean T1 values was assessed using the intraclass correlation coefficient (ICC), Pearson’s correlation coefficient and total least squares regression. In these analyses we treated the four ROIs on each image as independent measurements. The ICC describes the resemblance of two sets of data with identical units and an equal variance [20, 21], and can therefore be used to measure the agreement between the first and the registered second MR examination. A total least squares fit [22, 23] was performed to estimate a linear relation between measurements obtained at the two examinations. A fit with a slope significantly different from 1 would imply a systematic difference.

Results

Registration of T1-weighted images

Registration of the T1-weighted images improved the T1 mapping, as CRLBσ was reduced and the homogeneity of the T1 maps was increased (Fig. 2). Visual inspection of all ROIs showed an improvement in the alignment of the T1-weighted images due to registration. This effect was similar in the femoral and tibial cartilage ROIs. In only three out of 84 tibial cartilage ROIs, the severity of movement artefacts was increased by registration, which was caused by misalignment of the FSGPRTI=100 image in two of the cases.
Fig. 2

Comparison of T1 map and square root of CRLBσ map of a patient with and without automated registration to correct for patient movement (colour overlay). The CRLBσ (ms) provides a lower bound for standard deviation of T1 (ms) and is a measure for registration quality. The background greyscale image is a T1-weighted image (FSPGRTI = 2,100). a T1 map without registration, (b) CRLBσ map without registration showing high uncertainty in the T1 estimates, (c) improved T1 map with registration, (d) CRLBσ map with registration showing reduced uncertainty in the T1 estimates

Table 1 reports per subject the 90%-CRLBσ over all voxels in the eight ROIs, with and without registration. Table 2 shows the statistics for these measurements (first row) and for each ROI separately (rows 2–9). The 90%-CRLBσ decreased significantly in patients (P = 0.003) and controls (P = 0.005) owing to automated registration. The effect of registration was more pronounced in the femoral cartilage than in the tibial cartilage. With a Bonferroni correction for multiple testing, the registration effect in the femoral ROIs was still significant (P < 0.05).
Table 1

The 90%-CRLBσ calculated over all voxels in the eight cartilage ROIs

Subject

Patients

Controls

No registration

Automated registration

No registration

Automated registration

1

1,753

177

260

244

2

43

34

> 10,000

605

3

142

84

443

298

4

96

74

410

360

5

765

152

176

164

6

178

120

> 10,000

712

7

111

74

456

385

8

> 10,000

762

441

267

9

128

105

> 10,000

2,609

10

152

86

165

141

11

91

69

-

-

In one patient and three control subjects the 90%-CRLBσ value without registration is over 10,000 ms. This resulted from voxels where the maximum likelihood fit indicated an ‘infinitely’ long T1 time

Table 2

The effect of registration on the 90%-CRLBσ of all voxels in the eight ROIs combined and of each ROI separately

ROI

Patients (n = 11)

Controls (n = 10)

Median differencea

P valueb

Median difference

P value

All voxels in the eight ROIs

55.8

0.003

112.9

0.005

Femoral cartilage in the selected slice in the lateral compartment: Initial MRI

71.6

0.004

1,561.6

0.005

Femoral cartilage in the selected slice in the medial compartment: Initial MRI

223.2

0.003

210.9

0.005

Lateral femoral cartilage on second MRI

73.8

0.003

677.6

0.005

Medial femoral cartilage on second MRI

104.8

0.003

72.8

0.009

Tibial cartilage in the selected slice in the lateral compartment: Initial MRI

-5.7

0.374

-63.1

0.647

Tibial cartilage in the selected slice in the medial compartment: Initial MRI

-7.4

0.350

-5.1

0.508

Lateral tibial cartilage on second MRI

3.0

0.424

-45.0

0.959

Medial tibial cartilage on second MRI

-15.4

0.286

148.6

0.037

aDifferences in the median of the 90%-CRLBσ for all subjects between no registration and automated registration

bP values of the Wilcoxon signed rank test

Registration between the first and second MR examinations

Figure 3 shows scatter plots of the weighted mean T1 measurements on the first versus second MR examination. Table 3 summarises the statistics. In patients, the ICC for the manual slice-matching was 0.89 and with registration similar results were obtained (ICC = 0.90). In controls, an ICC of 0.85 was found with both manual matching and registration. These ICC values indicate a similar reproducibility for both methods as the ICC values lie in each other’s 95% confidence interval. The Pearson correlation coefficients show the same pattern. Linear total least squares regression resulted in a slope of approximately 1 in patients and controls, with both manual matching and automated registration, which confirms that there are no systematic differences.
Fig. 3

Weighted mean T1 value per ROI for all patients (a, b) and controls (c, d). Subjects are represented by different coloured markers. The weighted means of the T1 maps from the first and second MR examinations are plotted against each other with manual slice-matching (a, c) and automated registration (b, d). The black line represents y = x, which is the expected result at perfect registration and reproducibility; the blue line is a linear total least squares fit through the points; r represents the Pearson’s correlation coefficient of the points and ICC the intraclass correlation coefficient

Table 3

Comparison between first and second MR examinations: weighted mean T1 values of each ROI with manual slice-matching and automated registration

 

Patients

Controls

Manual matching

Registration

Manual matching

Registration

ICC(2,1) (CI)

0.893 (0.813–0.940)

0.902 (0.827–0.945)

0.849 (0.732–0.917)

0.851 (0.736–0.919)

r (CI)

0.892 (0.809–0.940)

0.901 (0.824–0.945)

0.855 (0.742–0.921)

0.858 (0.745–0.923)

slope (CI)

1.030 (0.412–1.649)

1.036 (0.383–1.688)

1.044 (0.589–1.499)

1.065 (0.506–1.624)

ICC(2,1) two-way random intraclass correlation coefficient based on single measures, r Pearson’s correlation coefficient, slope the slope obtained with a total least squares fit, CI 95% confidence interval

Discussion

Our study demonstrates that automated registration can improve two aspects of dGEMRIC studies in the setting of knee osteoarthritis: the T1 map calculation and the reproducibility of T1 maps obtained at different MR examinations.

First, applying automated image registration in the calculation of T1 maps from T1-weighted images improves the T1 maps and reduces the uncertainty in the estimated T1 value. In finger and hand arthritis, analysis by the coefficient of variation (CV) and visual grading analysis (VGA) showed that registration improves image quality and reduces variability [5]. In knee dGEMRIC, this has only been investigated in five healthy volunteers [6] where the goodness of the T1 fit was evaluated using χ2 analysis. In our paper the CRLBσ was used to measure the effect of registration on the T1 map. Like χ2 analysis, CRLBσ provides a quantitative measure to express the quality of the T1 fit, which is not provided by CV or VGA. However, χ2 and CRLBσ express different properties of the quality of fit: χ2 quantifies the difference between the fit and the measurements, and CRLBσ quantifies how noise on the MR signal propagates to the fit. CRLBσ has the advantage over χ2 of being an absolute measure expressed in the same units (ms) as T1. The results of our study show a reduction of CRLBσ due to registration, with a more pronounced effect in the femoral cartilage than in tibial cartilage. The 90%-CRLBσ did not show a significant effect of registration on the tibial ROIs (Table 2). This may be explained by the smaller signal intensity difference between cartilage and bone observed in the tibia compared with the femur. Misalignments in the tibial area will therefore have less effect on the CRLBσ than misalignments in the femoral region.

Second, automated image registration in evaluation of dGEMRIC reproducibility has been addressed. Using manual slice-matching, Multanen et al. [7] reported an ICC of 0.95 in femoral cartilage and 0.87 in tibial cartilage of control subjects and Siversson et al. [8] found a femur ICC of approximately 0.68 in patients. We found an ICC of 0.89 with the manual slice-matching method in patients. For the control group, the manual approach resulted in an ICC of 0.85, which is comparable to the aforementioned studies. The automated registration method presented in our study requires only one manually drawn cartilage mask for evaluating the same ROI in two or more MR examinations. Therefore, analysis of dGEMRIC is less time-consuming using the automated registration method. The actual amount of time that can be saved depends on the number of ROIs, slices and MR examinations to be analysed. This is particularly of interest if two or more MR examinations need to be analysed for large cohorts of patients. Automated registration yielded similar results to the manual method (ICC = 0.90 in patients and 0.85 in controls). The reproducibility in patients was slightly higher than in controls, which could be explained by differences in signal-to-noise ratio (SNR) owing to different coils, or by the difference in T1 values as controls underwent unenhanced MR imaging. The SNR was actually higher in controls than in patients (132 ± 38 vs 67 ± 37), so the better reproducibility in patients is most likely explained by the presence of contrast agent leading to shorter T1 values. The range of inversion times in the FSPGR sequence is optimised for these typical T1 values in the presence of contrast agent. The lower 90%-CRLBσ values for patients reflect this (Table 1).

Although this paper focuses on dGEMRIC, the registration method used is not specifically developed for dGEMRIC. The method should therefore be applicable to other MR mapping methods as well, such as T1ρ and T2.

In conclusion, automated registration of dGEMRIC in knee cartilage improves the quality of T1 maps and provides a good alternative to manual slice-matching in longitudinal studies as it leads to equal reproducibility and is operator-independent.

Notes

Acknowledgements

This project was funded by the Netherlands Organisation for Scientific Research (NWO).

References

  1. 1.
    Lawrence RC, Felson DT, Helmick CG et al (2008) Estimates of the prevalence of arthritis and other rheumatic conditions in the United States: Part II. Arthritis Rheum 58:26–35PubMedCrossRefGoogle Scholar
  2. 2.
    Bashir A, Gray ML, Burstein D (1996) Gd-DTPA2- as a measure of cartilage degradation. Mag Reson Med 36:665–673CrossRefGoogle Scholar
  3. 3.
    Bashir A, Gray ML, Boutin RD, Burstein D (1997) Glycosaminoglycan in articular cartilage: in vivo assessment with delayed Gd (DTPA)(2-)-enhanced MR imaging. Radiology 205:551–558PubMedGoogle Scholar
  4. 4.
    Bashir A, Gray M, Hartke J, Burstein D (1999) Nondestructive imaging of human cartilage glycosaminoglycan concentration by MRI. Magn Reson Med 41:857–865PubMedCrossRefGoogle Scholar
  5. 5.
    Miese F, Kröpil P, Ostendorf B et al (2011) Motion correction improves image quality of dGEMRIC in finger joints. Eur J Radiol 80:427–431Google Scholar
  6. 6.
    Studler U, White LM, Andreisek G, Luu S, Cheng HL, Sussman MS (2010) Impact of motion on T1 mapping acquired with inversion recovery fast spin echo and rapid spoiled gradient recalled-echo pulse sequences for delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) in volunteers. J Magn Reson Imaging 32:394–398PubMedCrossRefGoogle Scholar
  7. 7.
    Multanen J, Rauvala E, Lammentausta E et al (2009) Reproducibility of imaging human knee cartilage by delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) at 1.5 Tesla. Osteoarthritis Cartilage 17:559–564PubMedCrossRefGoogle Scholar
  8. 8.
    Siversson C, Tiderius CJ, Neuman P, Dahlberg L, Svensson J (2010) Repeatability of T1-quantification in dGEMRIC for three different acquisition techniques: two-dimensional inversion recovery, three-dimensional look locker, and three-dimensional variable flip angle. J Magn Reson Imaging 31:1203–1209PubMedCrossRefGoogle Scholar
  9. 9.
    Kellgren J, Lawrence J (1957) Radiological assessment of osteo-arthrosis. Ann Rheum Dis 16:494–502PubMedCrossRefGoogle Scholar
  10. 10.
    McKenzie CA, Williams A, Prasad PV, Burstein D (2006) Three-dimensional delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) at 1.5 T and 3.0 T. J Magn Reson Imaging 24:928–933PubMedCrossRefGoogle Scholar
  11. 11.
    Tiderius CJ, Olsson LE, Leander P, Ekberg O, Dahlberg L (2003) Delayed gadolinium-enhanced MRI of cartilage (dGEMRIC) in early knee osteoarthritis. Magnetic resonance in medicine 49:488–492PubMedCrossRefGoogle Scholar
  12. 12.
    Sijbers J, Den Dekker A, Raman E, Van Dyck D (1999) Parameter estimation from magnitude MR images. Int J Imag Syst Tech 10:109–114CrossRefGoogle Scholar
  13. 13.
    Cavassila S, Deval S, Huegen C, Van Ormondt D, Graveron-Demilly D (2001) Cramer-Rao bounds: an evaluation tool for quantitation. NMR Biomed 14:278–283PubMedCrossRefGoogle Scholar
  14. 14.
    Rao CR (1946) Minimum variance and the estimation of several parameters. Proc Cambridge Phil Soc 43:280–283CrossRefGoogle Scholar
  15. 15.
    Spandonis Y, Heese FP, Hall LD (2004) High resolution MRI relaxation measurements of water in the articular cartilage of the meniscectomized rat knee at 4.7 T. Magn Reson Imaging 22:943–951PubMedCrossRefGoogle Scholar
  16. 16.
    Li K, Zu Z, Xu J et al (2010) Optimized inversion recovery sequences for quantitative T1 and magnetization transfer imaging. Magn Reson Med 64:491–500PubMedGoogle Scholar
  17. 17.
    Poot DHJ, Den Dekker AJ, Achten E, Verhoye M, Sijbers J (2010) Optimal experimental design for diffusion kurtosis imaging. IEEE Trans Med Imaging 29:819–829PubMedCrossRefGoogle Scholar
  18. 18.
    Klein S, Staring M, Murphy K, Viergever MA, Pluim JPW (2010) Elastix: a toolbox for intensity-based medical image registration. IEEE Trans Med Imaging 29:196–205PubMedCrossRefGoogle Scholar
  19. 19.
    Klein S, Van der Heide UA, Lips IM, Van Vulpen M, Staring M, Pluim JPW (2008) Automatic segmentation of the prostate in 3D MR images by atlas matching using localized mutual information. Med Phys 35:1407–1417PubMedCrossRefGoogle Scholar
  20. 20.
    Shrout PE, Fleiss JL (1979) Intraclass correlations: uses in assessing rater reliability. Psychol Bull 86:420–428PubMedCrossRefGoogle Scholar
  21. 21.
    McGraw KO, Wong S (1996) Forming inferences about some intraclass correlation coefficients. Psychol Meth 1:30–46CrossRefGoogle Scholar
  22. 22.
    Petráš I, Bednárová D (2010) Total least squares approach to modeling: a Matlab toolbox. Acta Montanistica Slovaca 15:159–170Google Scholar
  23. 23.
    Golub GH, Loan CFV (1980) An analysis of the total least squares problem. SIAM J Num Anal 17:883–893CrossRefGoogle Scholar

Copyright information

© European Society of Radiology 2012

Authors and Affiliations

  • Esther E. Bron
    • 1
  • Jasper van Tiel
    • 2
    • 3
  • Henk Smit
    • 2
    • 4
  • Dirk H. J. Poot
    • 4
  • Wiro J. Niessen
    • 4
    • 5
  • Gabriel P. Krestin
    • 2
  • Harrie Weinans
    • 3
    • 6
  • Edwin H. G. Oei
    • 2
  • Gyula Kotek
    • 2
  • Stefan Klein
    • 4
  1. 1.Biomedical Imaging Group RotterdamDepartments of Medical Informatics and Radiology, Erasmus MCRotterdamThe Netherlands
  2. 2.Department of Radiology, Erasmus MCRotterdamThe Netherlands
  3. 3.Department of Orthopaedic Surgery, Erasmus MCRotterdamThe Netherlands
  4. 4.Biomedical Imaging Group Rotterdam, Departments of Medical Informatics and RadiologyErasmus MC, University Medical Centre RotterdamRotterdamThe Netherlands
  5. 5.Imaging Science & Technology, Department of Applied SciencesDelft University of TechnologyDelftThe Netherlands
  6. 6.Department of Biomechanical EngineeringDelft University of TechnologyDelftThe Netherlands
  7. 7.Department of Medical Informatics, Erasmus MCUniversity Medical Centre RotterdamRotterdamThe Netherlands

Personalised recommendations