M.G. Roberts, T.F. Cootes, E. Pacheco, J.E. Adams

1
Quantitative Vertebral Fracture Detection on DXA
Images using Shape and Appearance Models
M.G. Roberts, T.F. Cootes, E. Pacheco, J.E. Adams

Imaging Science and Biomedical Engineering,
University of Manchester, U.K.

2
Contents

• Clinical Background
• Appearance Models
• Classifier Training
• ROC curves
• Conclusions

3
Osteoporosis

• Disease characterised by
• Low bone mass and deterioration in trabecular
  structure
• Common Disease affects up to 40 of
  post-menopausal women
• Causes fractures of hip, vertebrae, wrist
• Vertebral Fractures
• Most common osteoporotic fracture
• Occur in younger patients, so provide early
  diagnosis

4
Classification
5
Limitations of current methods

• Morphometric Methods not reliable
• Use of 3 heights loses too much subtle shape
  information?
• No texture clues used (e.g. signs of collapsed
  endplate)
• But expert assessment has subjectivity problems
• Apparently widely varying fracture incidence
• Shortage of radiologists for expert assessment
• Availability of DXA Scanners in GP surgeries

6
Our Aims

• Automate the location of vertebrae
• Fit full contour (not just 6 points)
• Then use quantitative classifiers
• Use ALL shape information
• And texture around shape

7
DXA Images

• Very Low Radiation Dose
• Little or no projective effects
• Tilting Bean Can effects unusual
• Constant scaling across the image
• Whole spine on single image
• C-arms offer ease of patient positioning

8
Example Shape Fit
T12 wedge fracture
9
L2 Triplet Shape Modes 1-5
Derive shape models from manually annotated
training images
10
Appearance Models

• Combine Shape with Texture
• Sample image texture around/within shape
• Build texture model using PCA
• Combine shape and texture parameters
• Perform a tertiary PCA on combined vectors
• As shape/texture correlated
• This gives appearance model
• Appearance parameters determine both shape and
  texture

11
L2 Triplet Appearance Modes 1-3
12
Appearance Model Form

• Single vertebrae
• Models local edge structure in a region around
  the endplate

13
Classification Method

• Train Shape and Appearance Models
• Nearby Vertebrae are pooled
• T7-T9
• T10-T12
• L1-L4
• Refit Models to training images
• Record shape and appearance model parameters
• With fracture status
• Hence train linear discriminants
• Tried both shape and appearance parameters
• Used 3 standard height ratios as baseline
  comparison

14
Dataset

• 360 DXA Images
• 343 Fractures
• 97 Mild (Grade 1)
• 141 Moderate (Grade 2)
• 105 Severe (Grade 3)
• 187 non-fracture deformities
• Classified using ABQ method
• 2 radiologist consensus

15
Lumbar Spine ROC curves
16
T10-T12 ROC curves
17
T7-T9 ROC Curves
18
Grade 1 Fractures Combined
19
Grade 2 Fractures
20
FPR at 95 sensitivity
21
FPR on Grade 1 Fractures at 85 sensitivity
22
Conclusions

• Reliable quantitative classification on
  appearance model parameters
• 92 specificity at 95 sensitivity
• vs 79 specificity for standard morphometry
• Potential for clinical diagnosis tool (CAD)
• And use in clinical trials

23
For more information martin.roberts_at_manchester.ac
.uk www.isbe.man.ac.uk/mgr/autospine.html
This work was funded by the UKs ARC (Arthritis
Research Campaign) Earlier model development work
was funded by a grant from the Central Manchester
and Manchester Childrens University Hospitals
NHS Endowment Trust.
24
DIVA Tool
Whole spine view
Morphometry table classification Zoom view
User initialises solution by clicking on
approximate centres of vertebrae Then the tool
uses Active Appearance Model search to locate
shape contours around each vertebra