Automatic segmentation of lumbar vertebrae on digitised radiographs using linked active appearance models

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

Proc. Medical Image Understanding and Analysis, 2006. Vol. 1, pp.120-124.

Manual point placement for vertebral morphometry is time-consuming and imprecise, and morphometric methods for vertebral fracture diagnosis are unreliable. Automatic computer determination of the detailed vertebral shape could enable more powerful quantitative classifiers of osteoporotic vertebral fracture. The shape and appearance of vertebrae on 250 digitised lumbar radiographs were statistically modelled, using a sequence of active appearance models (AAMs) of overlapping triplets of vertebrae. To automatically locate the vertebrae, the sequence of models was matched to previously unseen scans. Accuracy results (0.64mm mean point-to-line error) were found to be similar to previously published results for dual-energy X-ray absorptiometry (DXA), but a low fracture prevalence meant that the shape models were undertrained for the few moderate and severe fractures. However mild fractures were fitted with good accuracy (mean 0.84mm). The results confirm the feasibility of substantially automating vertebral morphometry measurements on radiographs, despite the projective effects of the divergent X-ray beam. Use of the shape and appearance parameters of the models could in future provide a quantified form of some of the more subtle aspects of visual or semi-quantitative expert reading of vertebral fractures.