Search

Despite the advancement in imaging technologies, a fifth of the injuries in the cervical spine remain unnoticed in the X-ray radiological exam. About a two-third of the subjects with unnoticed injuries suffer tragic consequences. Based on the success of computer-aided systems in several medical image modalities to enhance clinical interpretation, we have proposed a fully automatic image analysis framework for cervical vertebrae in X-ray images. The framework takes an X-ray image as input and highlights different vertebral features at the output. To the best of our knowledge, this is the first fully automatic system in the literature for the analysis of the cervical vertebrae. The complete framework has been built by cascading specialized modules, each of which addresses a specific computer vision problem. This dissertation explores data-driven supervised machine learning solutions to these problems. Given an input X-ray image, the first module localizes the spinal region. The second module predicts vertebral centers from the spinal region which are then used to generate vertebral image patches. These patches are then passed through machine learning modules that detect vertebral corners, highlight vertebral boundaries, segment vertebral body and predict vertebral shapes. In the process of building the complete framework, we have proposed and compared different solutions to the problems addressed by each of the modules. A novel region-aware dense classification deep neural network has been proposed for the first module to address the spine localization problem. The proposed network outperformed the standard dense classification network and random forestbased methods. Location of the vertebral centers and corners vary based on human interpretation and thus are better represented by probability maps than single points. To learn the mapping between the vertebral image patches and the probability maps, a novel neural network capable of predicting a spatially distributed probabilistic distribution has been proposed. The network achieved expert-level performance in localizing vertebral centers and outperform the Harris corner detector and Hough forest-based methods for corner localization. The proposed network has also shown its capability for detecting vertebral boundaries and produced visually better results than the dense classification network-based boundary detectors. Segmentation of the vertebral body is a crucial part of the proposed framework. A new shapeaware loss function has been proposed for training a segmentation network to encourage prediction of vertebra-like structures. The segmentation performance improved significantly, however, the pixel-wise nature of proposed loss function was not able to constrain the predictions adequately. To solve the problem a novel neural network was proposed which predicts vertebral shapes and trains on a loss function defined in the shape space. The proposed shape predictor network was capable of learning better topological information about the vertebra than the shape-aware segmentation network. The methods proposed in this dissertation have been trained and tested on a challenging dataset of X-ray images collected from medical emergency rooms. The proposed, first-of-its-kind, fully automatic framework produces state-of-the-art results both quantitatively and qualitatively.