Caron Parsons1,2, Charles Hutchinson1,2, Emma Helm2, Alexander Kenneth Clarke3, Asfand Baig Mirza3, Qiang Zhang4, and Abhir Bhalerao4
1Division of Health Sciences, University of Warwick, Coventry, United Kingdom, 2Department of Radiology, University Hospital Coventry & Warwickshire, Coventry, United Kingdom, 3Warwick Medical School, Coventry, United Kingdom, 4Department of Computer Sciences, University of Warwick, Coventry, United Kingdom
Synopsis
There are multiple methods available for
skeletal age determination in the paediatric endocrine population. Only two
methods, using left hand and wrist x-rays are in frequent clinical use, however
Greulich & Pyle is based on data collated between 1931 and 1942 and Tanner
Whitehouse uses data from as far back as 1949. We present the initial results
of an automated software model of shape and textural analysis of the physes of
the knee. Introduction
Disorders
of growth and metabolism are a significant public health problem, responsible
for the majority of paediatric endocrinology referrals and a significant number
of consultations with general practitioners. Correlation of skeletal age (SA)
and chronological age (CA) alongside other clinical findings is key to the
diagnosis and management of many endocrine conditions, including puberty
disorders and short stature. These children often undergo serial x-rays during
their period of treatment to assess bone age. Recent studies of Japanese [1] and Italian [2] children examined the potential of
magnetic resonance images (MRI) of the left wrist for skeletal age estimation,
demonstrating correlation (R2 >0.9) between CA and SA.
Other
methods of SA determination include evaluation of knee x-rays [3], digital atlases of the wrist [4] and automated software models of
wrist x-rays [5], however to our knowledge there has
not been any evaluation of the textural and shape change at the physes on
MRI.
Advantages
of assessing the physes of the knee as a measure of skeletal maturity include
the fact that the knee is relatively easy to immobilize and image on MRI, and the
fact that the physes are large compared with those in small joints, therefore
providing a larger area over which to assess shape and texture. The relative
frequency with which knees are imaged clinically on MRI (compared with other
joints) gave us a substantial retrospective dataset with which to train and
test our algorithm.
Purpose
1) To train
and test a software model of MRI knee physis shape and texture for estimation
of SA.
2)
To compare the model's SA estimation with CA.
Methods
A
retrospective review of all paediatric MRIs of the knee performed at University
Hospital Coventry & Warwickshire identified 143 patients on the radiology
information system between 18/08/2010 and 10/08/2015. 4 cases were
excluded due to synovial abnormality or arthropathy and 6 were excluded due to
movement artefact or inadequate sequences.
The image
analysis consisted of three stages: (1) expert mark-up of training
images to build the shape and appearance models of variation [6]; (2) validation of the accuracy of
the models using leave-one-out testing on the training samples; (3) regression
of shape and texture co-factors from training samples to CA.
During
stage (1) we used proton-density and fat saturation sequences to produce
accurate surface delineations of the epiphyses (fig 1). The point data were converted
to a surface mesh, co-registered and the surfaces re-sampled with a smaller
sub-set of corresponding points at which appearance data was extracted. At
stage (2), we partitioned the training data and used cross-validation testing
to verify the accuracy of the shape and appearance model fitting. In addition,
the model characterised the textural changes across the epiphyseal plate and
this was used to produce an model of normal age-related change by using multivariate,
non-linear regression to CA. This was carried out by means of a neural network
classifier [7].
Results
Chronological age ranged between 4.99 and 18.4 years, of
which 43.4% were male. Overall root-mean-square error (RMSE) was 592 days (1.62
years) and the average absolute age prediction error was 457 days (1.25
years). Regression of the appearance and shape factors to CA showed
moderate correlation (R
2 = 0.634) (fig 2). Analysis of the degree of
error by age group demonstrated a higher mean error in the groups with smaller
samples (fig 3).
Discussion
SA determination is an important tool in paediatric
endocrinology, but assessment still relies on out-dated atlases [8, 9]
and the reader reliability of the interpreter, with intra-observer differences
of up to 0.96 years [10].
There are differences in SA versus CA that exist within different populations such
as ethnic background [11]
and obesity [12].
This work represents a first step in the development of an
automated software model for SA determination using knee MRI. Despite the small sample size, the average age
prediction error of 1.25 years lies within the 95% confidence intervals of
reported intra-observer error based on both the TW2 and GP atlases [10],
however relatively high age prediction errors occurred where there were smaller
numbers within an age group.
Conclusion
We present
the initial results of an automated software model of shape and textural
analysis of the epiphyses on knee MRI with a mean age prediction of 1.25 years.
Further normative data is needed to refine the software model., so that it can
be used for prediction of epiphyseal fusion and skeletal age determination.
Acknowledgements
No acknowledgement found.References
1. Terada,
Y., et al., Skeletal age assessment in
children using an open compact MRI system. Magn Reson Med, 2013. 69(6): p. 1697-702.
2. Tomei,
E., et al., Value of MRI of the hand and
the wrist in evaluation of bone age: Preliminary results. Journal of
Magnetic Resonance Imaging, 2014. 39(5):
p. 1198-1205.
3. O'Connor,
J.E., et al., A method to establish the
relationship between chronological age and stage of union from radiographic
assessment of epiphyseal fusion at the knee: an Irish population study. J
Anat, 2008. 212(2): p. 198-209.
4. Gilsanz,
V. and O. Ratib, Hand Bone Age. 2nd
ed. 2011: Springer.
5. Thodberg,
H.H., et al., The BoneXpert method for automated
determination of skeletal maturity. IEEE Trans Med Imaging, 2009. 28(1): p. 52-66.
6. Cootes,
T.F., G.J. Edwards, and C.J. Taylor, Active
appearance models, in Computer Vision
— ECCV’98, H. Burkhardt and B. Neumann, Editors. 1998, Springer Berlin Heidelberg.
p. 484-498.
7. Duda,
R.O., P.E. Hart, and D.G. Stork, Pattern
Classification (2nd Edition). 2000: Wiley-Interscience.
8. Greulich,
W. and P. Pyle, Radiographic atlas of
skeletal development of the hand and wrist. . 2nd ed. 1959, Stanford:
Stanford University Press.
9. Tanner
JM, W.R., Cameron N, Marshall WA, Healy M Jr, Goldstein H Assessment of skeletal maturity and prediction of adult height (TW2
method). 1983, London: Academic Press.
10. King,
D.G., et al., Reproducibility of bone
ages when performed by radiology registrars: an audit of Tanner and Whitehouse
II versus Greulich and Pyle methods. Br J Radiol, 1994. 67(801): p. 848-51.
11. Mora,
S., et al., Skeletal age determinations
in children of European and African descent: applicability of the Greulich and
Pyle standards. Pediatr Res, 2001. 50(5):
p. 624-8.
12. Johnson,
W., et al., Patterns of linear growth and
skeletal maturation from birth to 18 years of age in overweight young adults.
Int J Obes (Lond), 2012. 36(4): p.
535-41.