Quantitative Vertebral Fracture Detection on DXA Images using

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.

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

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

Classification

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

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

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

Example Shape Fit T 12 wedge fracture

L 2 Triplet Shape Modes 1 -5 Derive shape models from manually annotated training images

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

L 2 Triplet Appearance Modes 1 -3

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

Classification Method • Train Shape and Appearance Models • Nearby Vertebrae are pooled – T 7 -T 9 – T 10 -T 12 – L 1 -L 4 • 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

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

Lumbar Spine ROC curves

T 10 -T 12 ROC curves

T 7 -T 9 ROC Curves

Grade 1 Fractures Combined

Grade 2 Fractures

FPR at 95% sensitivity

FPR on Grade 1 Fractures at 85% sensitivity

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

For more information: martin. roberts@manchester. ac. uk www. isbe. man. ac. uk/~mgr/autospine. html This work was funded by the UK’s ARC (Arthritis Research Campaign) Earlier model development work was funded by a grant from the Central Manchester and Manchester Children’s University Hospitals NHS Endowment Trust.

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