Semi-automatic determination of detailed vertebral shape from lumbar radiographs using active appearance models

Abstract

Summary

The vertebral endplates in lumbar radiographs were located by a semi-automatic annotation method using statistical shape models.

Introduction

Vertebral fractures are common osteoporotic fractures, but current quantitative detection methods (morphometry) lack specificity. We have previously developed more specific quantitative classifiers of vertebral fracture using shape and appearance models. This method has only been applied to DXA vertebral fracture assessment (VFA) images and not to spinal radiographs. The classifiers require a detailed annotation of the outline of the vertebral endplate, so we investigated the application of similar semi-automated annotation methods to lumbar radiographs as the initial step in the generalisation of the statistical classifiers used in VFA images.

Methods

The vertebral body outlines in a training set of 670 lumbar radiographs were manually annotated by expert radiologists. This training set was used to build statistical models of vertebral shape and appearance using triplets of vertebrae. In order to segment vertebrae, the models were refitted using a sequence of active appearance models of vertebral triplets, using a miss-40-out train/test cross-validation experiment. The accuracy was evaluated against the manual annotation and analysed by fracture grade.

Results

Good accuracy was obtained for normal vertebrae (0.82 mm) and fracture grades 1 and 2 (1.19 mm), but the localisation accuracy deteriorated for grade 3 fractures to 2.12 mm.

Conclusion

Vertebral body shape annotation can be substantially automated on lumbar radiographs. However, an occasional manual correction may be required, typically with either severe fractures, or when there is a high degree of projectional tilting or scoliosis. The located detailed shapes may enable the development of more powerful quantitative classifiers of osteoporotic vertebral fracture.

Appendix 1—training set annotation tool

The images used for training were annotated manually using an in-house tool. The tool partly used previous AAMs to reduce the manual labour, with the models being iteratively updated as the annotated cases were added to the dataset. The current AAM set can be used to approximately locate the vertebrae. Then, the user moves any points which are not adequately located. A useful feature of this tool is that when points have been manually repositioned, they constrain other points in the model; so, if the initial automatic fit is incorrect, the user can partially correct it and then re-run the model fit as constrained by the corrections. For example, when four corner points and four mid-points (inner and outer superior and inferior rims) of a vertebra have been manually positioned the partially trained current AAM will typically produce good convergence on the rest of the vertebral shape in most cases, although fractured vertebrae may well need a few more points to be manually positioned due to their more complex shape.

There can be vertebrae (e.g. severe fractures) to which the provisional shape model cannot be fitted, due to undertraining. Therefore, we also used a dynamic programming edge search method in the tool. This first fits the vertebra's shape model learned so far to points already positioned by the user (e.g. the 4 corner points). Typically, this will not fit precisely due to undertraining, so next, the whole shape is warped using thin plate splines so that the shape exactly passes through the user-determined points. Then, those points not fixed by the user are moved towards the strongest local edge within 3 mm, but with high curvature penalised, and the optimum shape for the best combination of edge and curvature penalty costs is determined by dynamic programming (DP). If necessary, the user can further constrain the solution by fixing more points, or the user can accept the warped model fit prior to the DP edge–fit stage (e.g. in regions where there is no strong edge).

Hence, the tool is not constrained by the prototype models in cases where there is an inadequate fit; but as the annotated set is built up, it is possible to use the existing shape and appearance statistics to speed up annotating further vertebrae. Each time a new batch of images was annotated, all models were re-built. In the early stages, prototype shape models were created based on existing DXA VFA models, with a priori estimates of projective tilting effects using Bezier splines for the additional endplate rims. In the earlier batches of images (e.g. first 100 images), the AAMs were not used, and instead, the DP edge–fit method was used, and it was necessary to manually position a much higher proportion of points than in the later stages in which the models were becoming better trained.