1. To assess the accuracy of automatically located vertebral shapes on a challenging dataset of DXA images of the spine, including fractured vertebrae, and noisy images. 2. To improve the accuracy and robustness of the techniques.

A dataset of 202 DXA images of the spine was obtained, containing 173 fractures in total. The vertebral body outlines were annotated manually from L4 to T7, using 40 points each. Statistical models of vertebral shape and texture were derived, with the spine modelled by a sequence of overlapping vertebral triplets. A provisional texture model was derived on the best 120 images, and the extracted texture on the remaining (often very noisy) images was constrained by first fitting it to the provisional model. This reduces the number of clutter modes in the texture model, and improves overall robustness. The search algorithm fits the best sequence of models to an unseen image, using an Active Appearance Model for each vertebral triplet. Multiple triplet candidates are tried, but the best fitting one only is imposed at each iteration. Thus the ordering of the sequence is dynamically adjusted to defer poorly fitting regions until they are constrained by neighbours. Miss-N out tests were run to obtain estimates of accuracy, comparing the automatic and manual shapes.

A mean point-to-line accuracy of 0.7mm was obtained on normal vertebrae, well within the limits of manual precision. Over 95% of points in normal vertebrae were located with an accuracy better than 2mm. The corresponding accuracies for fractured vertebrae were: mean=1.2mm,median = 0.7mm, and 84% of points better than 2mm. Significant accuracy improvements were made by using the restricted texture model training, and by dynamically adjusting the order of the fit. However the models did not always match the more severe fractures, due to undertraining, and some confusion with the edges of neighbours.

Our latest results confirm the feasibility of substantially automating DXA vertebral morphometry even with fractures or noisy images. This should also enable the development of more reliable methods of fracture classification using the full shape and surrounding texture.