Quantitative Vertebral Fracture Dectection on DXA Images using

Quantitative Vertebral Fracture Dectection on DXA Images using Shape and Appearance Parameters Roberts MG, Cootes TF, Pacheco E, Adams JE University of Manchester, UK Results Introduction 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. Spinal Region Appearance Classifier FPR (%) 3 -height morphometry FPR (%) T 7 -T 9 3. 2% 21. 6% T 10 -T 12 4. 7% 18. 5% L 1 -L 4 4. 9% 10. 0% False positive rates (FPR) (%) at 95% sensitivity for an appearance classifier compared with standard 3 height morphometry b a Fig a) shows the lumbar part of a DXA image, with the detailed segmentation curves superimposed. Fig b) shows a zoomed-in view of the points that are used to produce the shape models. T 9 (severe fracture) and T 10 are shown. There are 32 points per vertebral body, and pedicle connections for T 10 and below. 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. Material & Methods A fracture-rich dataset of 360 DXA images was used. The vertebral body outlines were manually annotated from L 4 to T 7. 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.