| dc.description.abstract | Medical imaging techniques, such as multi-modality Magnetic Resonance Imaging (MR) and Computed Tomography (CT), have been invaluable for the early detection, diagnosis, and treatment of diseases. The process of annotating large-scale medical datasets is labor-intensive and generally a prerequisite for the employment of deep learning technologies. Despite complexity, the exceptional segmentation capabilities of deep learning, ranging from brain to whole-body imaging, cannot be ignored. This dissertation focuses on characterizing the connections between the brain and body through data-efficient segmentation techniques. The research is methodically divided into two main sections: brain and body. In the brain section, a 4D white matter atlas is initially constructed on a standard template. Following this, a spatially localized deep neural network is proposed for the direct segmentation of white matter from structural MRI. In the body section, a two-stage transfer learning method is proposed to quantify body composition on CT mid-thigh slices. To enhance muscle anatomy understanding, domain adaptation and self-training are leveraged to transfer labels from public MRI to individual CT slices. In the final chapter, we characterize relationship between brain features, such as brain region volume, and body features like body composition, by predicting each set of features from the other. In alignment with this objective, the Gumbel-softmax method is utilized to select informative features for the enhanced prediction of brain and body features reciprocally. The dissertation concludes by outlining the significant contributions of this research and proposing future work. | |
