Resolution Independent Visualization for Comparative Analysis of Multi-Dimensional Imaging Data
Dr Yu, Yizhou (Principal investigator)
Visualization, Geometry Processing, Data Analytics
Others - Computing Science and Information Technology
Block Grant Earmarked for Research (104)
HKU Project Code
Seed Funding Programme for Basic Research
A time of opportunities is also a time of challenges. To enable unprecedented capabilities in medical image analytics, for instance, the discovery of patterns, relationships and abnormalities within and across populations, there is a pressing need for a systematic integration and coordinated visualization framework and their associated data modeling and mining technologies for the large amounts of multi-measure multi-dimensional medical imaging data. In many types of neurological disorders, such as epilepsy and Tourette Syndrome (TS), it is very hard to localize pathological brain tissues because the brain is a highly complex functional unit still with many unsolved mysteries. Neurologists and neurosurgeons need to refer to existing normal and abnormal cases before making a diagnosis. Existing abnormal cases with similar brain pathologies are particularly helpful in suggesting which measurements and symptoms to look at. In current practices, neurologists still need to manually initiate data processing for one subject at a time and then prepare spreadsheets or slides to compare the results with reference cases side by side. To make this comparative analysis possible, they need to empirically speculate beforehand which reference cases, measurements and 2D image slices are likely to be the most relevant ones. If their speculation fails, they have to repeat the whole process on another subset of reference cases, measurements and/or image slices until they find the right ones. The process of conducting a single pairwise full comparison usually takes them half a day. The labor-intensive and time-consuming nature of the procedure makes it impossible to fully take advantage of the large number of imaging datasets available for reference and comparison, and reach more informed and accurate decisions. The overall objective of the proposed research is to design and develop a rigorous framework based on advanced data abstraction and geometric computing for efficient and effective management, exploration, comparison and visualization of large collections of multi-measure multi-dimensional medical images. ‧ Design and develop a coordinated visualization framework for comparative analysis of multi-measure multi-dimensional medical images. Such a framework includes both data abstraction and visual abstraction. We propose to adopt advanced data abstraction methods and perform deep analysis to abstract and aggregate rich spatially varying measurements inside multi-dimensional medical images as well as semantic regions. We propose to simultaneously visualize multiple images with multiple measures in a table-based layout to facilitate side-by-side comparative analysis. Furthermore, every pane in the table is capable of showing a distinct 3D interactive medical visualization instead of 2D image slices. We plan to achieve coordinated visualization, exploration and comparison of the organized data across multiple panes using synchronized camera movements, similarity-based ranking and visual enhancement of regions with salient differences. We also propose multiple strategies, including LOD representations and GPU acceleration, to overcome performance hurdles. ‧ Design a novel multi-scale framework for multi-dimensional image segmentation, which subsequently enables multi-dimensional image feature extraction and vectorization. Coordinated visualization and exploration of a collection of datasets requires accurate spatial segmentation of the datasets. Apparently, intrinsic geometric data embedded in 3D imaging of real-world objects is very important in linking individual objects for interpretation. We propose to process data extracted from 3D images in a novel geometry scale space through the integration of level sets and scale space theory. The proposed multi-scale segmentation framework in the geometry scale space can support accurate representation and comparison related to shape similarity. Subsequently, measurements within segments can be computed supporting coordinated visualization and comparative analysis. ‧ Develop powerful resolution independent vector representations for high-definition visualization of multi-dimensional images. Compare to line drawings, such images need a more powerful vector representation that accounts for color variations across the image space in addition to curvilinear features. In addition to developing novel and powerful resolution independent representations, other key issues include vectorization and rasterization algorithms that can convert raster 2D and volumetric images to their corresponding vector representations and vice versa. The vectorization algorithm includes a mechanism for detecting and embedding multi-scale sharp features as well as methods for representing both low-frequency and high-frequency color variations within image regions. The rasterization algorithm needs to display vectorized images in a responsive way.