Structural Medical Image Analyses using Consistent Volume and Surface Image Processing
Medical imaging refers to the technologies of creating visual representation of the interior of human body. Clinical practitioners can make diagnoses by visually investigating the qualitative medical images, which relies on the experts’ experiences. In past decades, medical image analysis algorithms have been developed to obtain quantitative information from medical images. Historically, the medical image analysis on structural images was limited to a small-scale cohort (e.g., <500 images), whose images were collected from a single scanner. Recent developments on data sharing and computational power offer us an opportunity to explore large-scale medical image data. In this dissertation, we present large-scale medical image processing and analyses methods for both brain and abdomen. For the brain, we have established an end-to-end large-scale medical image analysis framework in investigating lifespan aging by conducting robust and consistent whole brain volume and surface metrics, controlling inter-subject variations, and conducting robust statistical analyses. We have generalized the multi-atlas label fusion theory from 3D to 4D for longitudinal whole brain segmentation. For the abdomen, we have proposed splenomegaly segmentation methods using multi-atlas approach, deep convolutional neural networks, and synthesis learning. Then, we applied abdomen segmentation methods to characterize 3D structure of the pyelocalyceal anatomy.