Artificial neural networks for the recognition of vertebral landmarks in the lumbar spine

The diagnosis and treatment of spinal disorders often requires the measurements of anatomical parameters on radiographic projections, which is usually performed manually. Due to the non negligible degree of intra- and inter-observer variability of these measurements, a user-independent, automated method for the characterisation of the spinal anatomy is needed. Artificial neural networks are currently used for many automated tasks in which a robust, fault-tolerant performance is needed, and may prove to be useful for this task. In this paper, a novel method based on a neural network aimed to the automatic identification of vertebral landmarks is presented. A radiographic database of lumbar sagittal radiographic projections vertebrae of adult patients suffering from various spinal disorders was created. Vertebral landmarks at the projected corners of the vertebral endplates of L3 and L4 were manually identified in all images. The annotated images were used to train and test an artificial neural network in the automatic recognition of such landmarks. The values of clinically relevant anatomical parameters (disc and vertebral heights, disc wedging) were then geometrically calculated based on the predicted landmark coordinates and compared to manual measurements.The novel method proved to be able to identify vertebral landmarks, with errors and limitations which should be taken into account. Possible future applications of neural network-based methods include the automatic extraction of clinically relevant parameters from radiographic images of the lumbar spine.