P.A. Bromiley, J. Adams and T.F. Cootes
In: Journal of Orthopaedic Translation Volume 2 Issue 4: 20th International Bone Densitometry
Workshop (IBDW); 13 Oct 2014-17 Oct 2014; Hong Kong. 2014. p. 227-228.
Osteoporotic fractures are associated with significant morbidity, mortality and public
health costs, and will increase with an ageing population. Many osteoporotic vertebral
fractures (VF) present on images do not come to clinical attention or lead to fracture
prevention treatment. Furthermore, DXA vertebral fracture assessments (VFA) are often
reported subjectively by a radiologist or other clinician. VFA computer-aided systems
offer potential advantages. Methods based on statistical shape models (e.g. active
appearance models, AAMs) have been used to segment vertebrae in radiographs and DXA
VFA. However, results achieved using AAMs exhibit significant numbers of large errors
due to model fitting failure, particularly on more severely fractured vertebrae.
We evaluate an alternative algorithm, the Random Forest Regression Voting Constrained
Local Model (RFRV-CLMs), which has proved more robust and generalizable than AAMs
in annotation of landmarks on various clinical images; we investigate whether this
will reduce the number of fitting failures in vertebral segmentation.320 DXA VFA images
obtained on various Hologic (Bedford MA) scanners had manual annotations of 405 landmark
points of vertebrae T7 to L4, with fracture classifications from an expert radiologist.
RFRV-CLMs were applied to these data in a leave-1/4-out fashion. Figure 1 shows an
example image with manual annotations (left; grade 3 severe and grade 1 mild fractures
of L1 and L2, respectively), with automatic annotations for L1 to L3 (right). Errors
were calculated as the mean, across each vertebra, of the minimum distances between
the automatic annotations and a Bezier spline passing through the manual annotations.
Figure 2 shows a cumulative distribution function of errors in ten vertebral levels
in all 320 images, for each vertebral classification. Mean errors of \textless4mm
were achieved for 95\% of grade 3 fractures, and 100\% of other classifications.
Mean errors above 2mm were observed on 3.84\% and 10.48\% of vertebra with grade 2
and 3 fractures, respectively, compared to 10.2\% and 16.5\% obtained with AAMs.
We conclude that the RFRV-CLM produces fewer large errors due to fitting failures
than the AAM in vertebral segmentation, and so is more suitable for automatic analysis
of DXA VFA.