Osteoporotic vertebral fractures (VFs) are under-diagnosed, creating an opportunity for computer-aided, opportunistic fracture identification in clinical images. VF diagnosis and grading in clinical practice involves comparisons of vertebral body heights. However, machine vision systems can provide a high-resolution segmentation of the vertebrae and fully characterise their shape and appearance, potentially allowing improved diagnostic accuracy. We compare approaches based on vertebral heights to shape/appearance modelling combined with k-nearest neighbours and random forest (RF) classifiers, on both dual-energy Xray absorptiometry images and computed tomography image volumes. We demonstrate that the combination of RF classifiers and appearancemodelling, which is novel in this application, results in a significant (up to 60% reduction in false positive rate at 80% sensitivity) improvement in diagnostic accuracy.
展开▼
