| dc.description.abstract | The abdominal organs, such as liver, spleen, pancreas, and kidneys are crucial organ systems and the relationships between spatial form and function are important area of study. Computed tomography (CT) of the abdomen is an essential clinical tool in diagnostic investigation and efficient quantitative measurement for internal organs, bones, soft tissue, blood vessels, and it allows for identification of structures in possible abnormalities and tumors. This dissertation develops techniques to study the abdominal organs and body compositions through volumetric CT. Targeting discoveries of meaningful quantitative biomarkers, we create the data-driven artificial intelligent (AI) systems for clinically acquired medical images. While the work can be generalized to other field of studies such as image-based recognition and perceptions, we focus on the techniques for spatial anatomical context from imaging data. This dissertation proposes methods motivated by large-scale heterogeneous radiography images and focuses on medical image analysis techniques to best exploit and explore information in the human body.
In this research, we focus on 8 major contributions. First, we query and extract large-scale study cohorts from ImageVU. To curate and assure the research quality data, we investigate the body part regression given CT series studies. We propose to use the self-supervised learning and semi-supervised learning techniques for regularizing the CT slice orders and distances according to the anatomical landmarks. Second, we focus on the heterogeneous contrast of dynamic CT images, a GAN-based method is propose to learn a disentangled representation of the specified internally variation of five enhancement phases. Next, the studies of 3D high-resolution medical image segmentation is one of the most essential mean for quantifying the measurement of organs and tissues. We propose a coarse-to-fine solution named RandomPatch, which accurately segmented 13 organs in the abdomen. In order to incorporate the rich anatomical context in the segmentation technique, we then 1) create a new joint modeling method that combine the non-imaging clinical phenotypes and the CT images for pancreas segmentation; 2) propose to exploit the advantages of spatial long-range dependencies by transformer model, specifically, we build a hybrid model that take the 3D image as input for transformer encoder, then decodes by the CNN for its superiority of local context modeling; 3) are motivated by the self-supervised pre-training tasks of image inpaintng, contrastive learning and rotation prediction, we present the first hierarchical segmentation model pre-trained by large-scale unlabeled CT volumes; 4) target the efficiency and scalability of 3D segmentation models, we show the aggregation block for transformer layers have a better capability for robustness. After acquiring the robust quantification tools for interpreting 3D medical images, we present the first studies of abdominal atlases for accurate tissue correspondence. The pancreas, kidney and its substructure atlas template are constructed using registration and segmentation frameworks. Lastly, we demonstrate these automatic machine learning models can well measure the splenomegaly patients and provide quantitatively analysis for exploratory biomarkers. We conclude this dissertation by contributions, challenges, visions and future work, and trust the science and engineering technologies can make healthcare and research work innovative. | |
