Visualization of conserved structures by fusing highly variable datasets.

Abstract

INTRODUCTION: Skill, effort, and time are required to identify and visualize anatomic structures in three-dimensions from radiological data. Fundamentally, automating these processes requires a technique that uses symbolic information not in the dynamic range of the voxel data. We were developing such a technique based on mutual information for automatic multi-modality image fusion (MIAMI Fuse, University of Michigan). This system previously demonstrated facility at fusing one voxel dataset with integrated symbolic structure information to a CT dataset (different scale and resolution) from the same person. The next step of development of our technique was aimed at accommodating the variability of anatomy from patient to patient by using warping to fuse our standard dataset to arbitrary patient CT datasets. METHODS: A standard symbolic information dataset was created from the full color Visible Human Female by segmenting the liver parenchyma, portal veins, and hepatic veins and overwriting each set of voxels with a fixed color. Two arbitrarily selected patient CT scans of the abdomen were used for reference datasets. We used the warping functions in MIAMI Fuse to align the standard structure data to each patient scan. The key to successful fusion was the focused use of multiple warping control points that place themselves around the structure of interest automatically. The user assigns only a few initial control points to align the scans. Fusion 1 and 2 transformed the atlas with 27 points around the liver to CT1 and CT2 respectively. Fusion 3 transformed the atlas with 45 control points around the liver to CT1 and Fusion 4 transformed the atlas with 5 control points around the portal vein. The CT dataset is augmented with the transformed standard structure dataset, such that the warped structure masks are visualized in combination with the original patient dataset. This combined volume visualization is then rendered interactively in stereo on the ImmersaDesk in an immersive Virtual Reality (VR) environment. RESULTS: The accuracy of the fusions was determined qualitatively by comparing the transformed atlas overlaid on the appropriate CT. It was examined for where the transformed structure atlas was incorrectly overlaid (false positive) and where it was incorrectly not overlaid (false negative). According to this method, fusions 1 and 2 were correct roughly 50-75% of the time, while fusions 3 and 4 were correct roughly 75-100%. The CT dataset augmented with transformed dataset was viewed arbitrarily in user-centered perspective stereo taking advantage of features such as scaling, windowing and volumetric region of interest selection. CONCLUSIONS: This process of auto-coloring conserved structures in variable datasets is a step toward the goal of a broader, standardized automatic structure visualization method for radiological data. If successful it would permit identification, visualization or deletion of structures in radiological data by semi-automatically applying canonical structure information to the radiological data (not just processing and visualization of the data's intrinsic dynamic range). More sophisticated selection of control points and patterns of warping may allow for more accurate transforms, and thus advances in visualization, simulation, education, diagnostics, and treatment planning.