查看文章

Accurate segmentation of the spine in computed tomography (CT) images is mandatory for quantitative analysis, e.g. in osteoporosis, but remains challenging due to high variability in vertebral morphology and spinal anatomy among patients. Conventionally, spine segmentation was performed by model-based techniques employing spine atlases or statistical shape models. We argue that such approaches, even though intuitive, fail to address clinical abnormalities such as vertebral fractures, scoliosis, etc. We propose a novel deep learning-based method for segmenting the spine, which does not rely on any pre-defined shape model. We employ two networks: one for localisation and another for segmentation. Since a typical spine CT scan cannot be processed at once owing to its large dimensions, we find that both nets are essential to work towards a perfect segmentation. We evaluate our framework on three …