Visuelle Analyse in der Epidemiologie
Final Report Abstract
Interpreting dynamic data requires substantial domain expertise, as in all cases the data is rich but severely distorted by motion and noise. The main problem is that these artifacts share properties with the information that we want to analyze, since both produce a changing signal over time. Hence, information buried in the data needs to be extracted by a sequence of analysis steps where domain knowledge is used to remove or reduce the different kinds of adverse influences (noise, motion from breathing, patient motion and muscle relaxation) while preserving the time-varying signal. Our general strategy in the project so far was to use visual analytics strategies not only for the final feature analysis but also at earlier stages of feature generation to get a more general solution to understand this kind of data. Hence, we simultaneously worked on the preprocessing steps for generating features and on ways to select or extract features. In the example of grouping features by spatial and temporal similarity, we were able to show that steering the grouping process by evaluating the shape and distribution of feature carriers (the voxels in this case) allowed selecting an appropriate grouping strategy based on visual appearance of perfusion features. However, currently automatic selection, e.g. by using a high-level perfusion model, is still more effective since appearance of spatial arrangement in 3D is visualized only slice-wise. This was not true of course for the evaluation of temporal similarity alone as it was shown in the experiments using linear decorrelation to detect prominent signal while suppressing effects from noise. Spatial arrangement is important and we need to look more carefully into this because our work on correcting mis-registration although currently being completely automatic could be complemented by a visually-guided component if distortion from noise and motion artifacts is too high for allowing separation of features from artifacts simply based on low-level information. This is the case for the CEUS data. With regard to the application-oriented research questions, so far, the project was very successful, as in the evaluation by the domain scientist we could show that it is indeed possible to define a common analysis environment that can be adapted to different analysis tasks for dynamic data. Also, we were able to show that using visual analysis techniques allowed to evaluate dynamic data based on spatial arrangement in 3D (which has not been done before) and that visual analysis of time-dependent similarity constraints via PCA enabled us to efficiently separate feature information from noise in the data. While the advantages from integrating preprocessing and visual analytics in a closer way were as anticipated, we found it difficult to organize the work plan such that synergy really took place. Image analysis and pattern recognition aim at automating decision finding, and visual computing generally assumes that remaining distortions can be detected and corrected in an interactive manner. This leads to an expectation that methodology for preprocessing exists before visual analysis tools are investigated. Breaking these dependencies required more coordination work than we expected and also required to change the work plan to some extent. Since results show that efficiency in adapting analysis tools as well as in actually using them to carry out an analysis can be gained by this kind of integration, we intend to push this collaboration even further in the continuation project, but we also organized the work plan in a different way that better reflects the necessity that methodology development on both sides requires a certain amount of time.
Publications
- A Visual Analytics Approach to Diagnosis of Breast DCE-MRI Data, Proc. of Vision, Modeling, and Visualization (VMV) pp. 351–362, 2009
S. Glaßer, S. Schäfer, S. Oeltze, U. Preim, K.-D. Toennies, B. Preim
- Visual Analysis of Cerebral Perfusion Data – Four Interactive Approaches and a Comparison, Proc. of the 6th International Symposium on Image and Signal Processing and Analysis (ISPA); pp.588-595, 2009
S. Oeltze, H. Hauser, J. Rorvik, A. Lundervold, B. Preim
- A visual analytics approach to diagnosis of breast DCE-MRI data, Computer and Graphics, Vol. 34 (8):602-611, 2010
S. Glaßer, U. Preim, K.-D. Toennies, B. Preim
- Detection of Motion Distorted Areas in Perfusion MRI of the Breast, Bildverarbeitung für die Medizin (BVM2010), pp. 172–176, 2010
S. Schäfer, K.-D. Tönnies
- Local Similarity Measures for ROI- based Registration of DCE-MRI of the Breast, Proc. of Medical Image Understanding and Analysis (MIUA), pp. 159-163, 2010
S. Schäfer, C.M. Hentschke, K.-D. Tönnies
