Vertebral shape: automatic measurement by DXA using overlapping statistical models of appearance

1
Vertebral shape automatic measurement by DXA
using overlapping statistical models of
appearance

• Martin Roberts and Tim Cootes and Judith Adams
• martin.roberts_at_man.ac.uk

Imaging Science and Biomedical Engineering, Univer
sity of Manchester, UK
2
Contents

• Osteoporosis - Background
• DXA vs Conventional Radiography
• Fracture Classification
• Our aims in automating vertebral DXA
• Automatic Location Method
• Results for Vertebral Morphometry Accuracy
• Conclusions

3
Osteoporosis

• Disease characterised by
• Low bone mass or
• 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
Osteoporosis Vertebral Fractures

• A vertebral fracture indicates increased risk of
  future fractures
• the risk of a future hip fracture is doubled (or
  even tripled)
• the risk of any subsequent vertebral fracture
  increases five-fold
• A very important diagnosis for radiologists to
  make
• Incident vertebral fractures used in clinical
  trials
• To assess the efficacy of osteoporosis therapies

5
Osteoporosis - statistics

• 40 of middle-aged women in Europe affected
• 200,000 osteoporotic fractures per year in the UK
  
• Half of these are vertebral
• Given one vertebral fracture
• the risk of a future hip fracture is doubled (or
  even tripled)
• the risk of any subsequent vertebral fracture
  increases five-fold

6
Advantages of DXA

• Very Low Radiation Dose
• 1/100 of spinal radiographs
• Little or no projective effects
• Bean Can effects unusual
• Constant scaling across the image
• Whole spine on single image
• C-arms offer ease of patient positioning
• Convenient as supplement to BMD scan

7
Disadvantages of DXA

• Definition and resolution much poorer than
  conventional radiography
• But now improved to 0.35mm line pair
• Images suffer from high noise or clutter from
  ribs soft tissue
• Upper vertebrae (T4-T6) often poorly visualised
• But these have lower fracture incidence
• Arm positioning can also help

8
Example DXA image lateral view of spine
Disadvantages Very low dose but noisy Poorer
resolution than radiography (0.35mm vs
0.1mm) Above T7 shoulder-blades can cause poor
imaging of T6-T4
9
Comparisons with spinal radiography

• Good concordance between visual DXA and visual XR
  
• Ferrar et al JBMR 2003
• Good concordance between morphometric methods for
  DXA and Radiography
• Rea Ost Int 1999, Ferrar JBMR 2000
• Majority of discrepancy over Grade 1 mild
  fractures/deformities or T6-T4
• Useful pre-screen to avoid higher radiation
  spinal radiographs

10
(No Transcript)
11
Classification methods

• Quantitative morphometry - height ratios
• Much shape information discarded
• (3 heights)
• Texture clues unused
• e.g. wider texture band around an endplate
  collapse
• So visual XR or Genant semi-quantitative more
  favoured
• But subjectivity still a problem for mild
  fractures
• Mild deformities may be mis-classed as fractures
• Algorithm-based qualitative identification (ABQ)
• Comparison of methods for the visual
  identification of prevalent vertebral fracture in
  osteoporosis.Jiang G, Eastell R, Barrington NA,
  Ferrar L.Osteoporos Int. 2004 Apr

12
Our Aims

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

13
Automatic Location

• User clicks on bottom, top and middle vertebrae
• Start at mean shape through these 3 points
• Fit a sequence of linked appearance models
• Overlapping triplets
• E.g (L4/L3/L2), and (L3/L2/L1) etc
• Overlaps give helpful linking constraints
• Sequence Order is dynamically adjusted based on
  local quality of fit
• High noise or poor fit regions deferred

14
Appearance Models

• Statistical Model of both shape and surrounding
  texture
• Learned from a training set of manually annotated
  images
• Good robustness to noise
• shapes constrained by training set
• But need large training set to fit to extreme
  pathologies
• (e.g. grade 3 fractures)

15
Example AAM fit to DXA image
User initialises by clicking 3 points at bottom,
middle, top (L4, T12, T7).
16
Dataset

• 184 DXA images
• 80 images contain fractures
• 137 vertebral fractures
• Also a bias towards obese patients
• So often high noise in lumbar
• Some other pathologies present
• Disk disease, large osteophytes
• So challenging dataset

17
Experiments

• Repeated Miss-4-out tests
• 180 image Training Set and 4 Test Set partition
• 10 replications with emulated user-supplied
  initialisation (Gaussian errors)
• Manual annotations as Gold Standard
• Mean Abs Point-to-Curve Error per vertebra
• Percentage number of points within 2mm also
  calculated

18
Automatic Search Accuracy Results
Vertebra Status Median (mm) 90-ile (mm) Pts Errorlt2
Normal 0.73 1.20 98.2
Fractured or Deformed 0.94 2.82 84.6
Search Errors (per vertebra pooling T7-L4)
Some under-training for fractures causes long
tail
19
Conclusions

• Good automatic accuracy on normal vertebrae
• Promising accuracies on fractured vertebrae
• Need to extend training set
• Vertebral shapes can be reliably located on DXA
  with only minimal manual intervention
• This allows a new generation of quantitative
  classification methods
• Could extend to digitised radiographs

20
Acknowledgements

• Acknowledge assistance of
• Bone Metabolism Group, University of Sheffield
• R Eastell, L Ferrar, G Jiang

21
For more

FOR MORE INFO...

• www.isbe.man.ac.uk

• martin.roberts_at_man.ac.uk