Quantitative Vertebral Fracture Dectection on DXA Images using Shape and Appearance Parameters

1
Quantitative Vertebral Fracture Dectection on DXA
Images using Shape and Appearance Parameters
Roberts MG, Cootes TF, Pacheco E, Adams
JE University of Manchester, UK
Introduction
Results

• Vertebral fractures are a strong indicator of
  osteoporosis. Quantitative morphometric
  assessment for vertebral fracture can be made
  from DXA images of the spine, but is
  insufficiently specific.
• We used detailed statistical models of the shape
  and appearance of vertebrae to develop more
  reliable quantitative classification methods.

False positive rates (FPR) () at 95 sensitivity
for an appearance classifier compared with
standard 3-height morphometry

• At 95 sensitivity the appearance classifier has
  an overall false positive rate under 5, compared
  to 18 false positive rate with standard
  morphometry.
• This produces a sensitivity of 85 for grade 1
  fractures, compared to 65 using height-based
  morphometry.

b
Fig b) shows a zoomed-in view of the points that
are used to produce the shape models. T9 (severe
fracture) and T10 are shown. There are 32 points
per vertebral body, and pedicle connections for
T10 and below.
a
Fig a) shows the lumbar part of a DXA image, with
the detailed segmentation curves superimposed.
Material Methods

• A fracture-rich dataset of 360 DXA images was
  used. The vertebral body outlines were manually
  annotated from L4 to T7.
• Statistical models of vertebral shape and texture
  were derived from this annotated training set.
• The shape and texture model parameters were then
  combined to create appearance models for the
  lumbar, lower-thoracic and mid-thoracic
  vertebrae.
• The vertebrae were visually assessed by two
  radiologists using the ABQ method. A consensus
  was then reached for any vertebrae with
  discrepancies in classification. There were 354
  fractures and 158 other short vertebral height
  deformities identified.
• The shape and appearance models were re-fitted to
  each training image, and their resulting model
  parameters used to train linear discriminant
  classifiers, given the radiologists gold
  standard.
• Classifier performance on unseen data was
  assessed using leave-1-out train/test
  experiments, and ROC curves were derived by
  varying the fracture detection threshold.

Combined ROC curves for appearance and shape
classifiers, and 3-height morphometry
Conclusions

• The appearance classifer can distinguish between
  true vertebral fractures and other short
  vertebral height deformities more reliably than
  height-based methods.
• The appearance classifier can operate at around
  95 sensitivity and specificity on DXA data.
• This technique has the potential to be used
  clinically, when combined with Active Appearance
  Model segmentation.