HighLevel User Interfaces for Transfer Function Design with

High-Level User Interfaces for Transfer Function Design with Semantics Christof Rezk Salama (Univ. Siegen , Germany) Maik Keller (Univ. Siegen, Germany) Peter Kohlmann (TU Vienna, Austria) institute for vision and graphics university of siegen, germany

Volume Visualization Volume visualization techniques are mature from the technical point of view. Real-time volume graphics on commodity PC hardware Multidimensional transfer functions/classification Gradient estimation and local illumination on-the-fly Memory management and compression for large volumes Even global illumination techniques. Is the ”volume rendering problem“ solved? If you ask the computer scientist, he‘ll probably say „yes“. If you ask the users, they will most likely say „no“ christof rezk-salama, institute for vision and graphics, university of siegen

Questions Why are volume rendering applications so hard to use for nonexperts? Are volume rendering applications easy to use for us, the „experts“ ? What features must appropriate user interfaces provide? christof rezk-salama, institute for vision and graphics, university of siegen

The Mental Model Example taken from: Donald A. Norman The Psychology of Everyday Things christof rezk-salama, institute for vision and graphics, university of siegen

Volume Visualization Transfer Function Design: Mapping of scalar data to optical properties (emission/absorption) Color table: Example: 1 D TF for 12 bit Data, 4096 values x RGBA = 16384 DOF Editors based on geometric primitives 1 D Transfer Functions 2 D Transfer Functions christof rezk-salama, institute for vision and graphics, university of siegen

User Intention Examples: „Fade out the soft tissue“ „Sharpen the blood vessels“ „Enhance the contrast“ Question: What actions are necessary? Even the expert, who programmed the user interface, does not know this! Mental model is inappropriate or missing! Semantics are missing (leads to “gulf of execution”) Result in trial-and-error christof rezk-salama, institute for vision and graphics, university of siegen

Abstraction Levels Semantic Level Visibility Sharpness Contrast All previous approaches aim at reducing the complexity, the degrees of freedom. None of the prevous approaches tries to provide an appropriate mental model! User High-Level Parameters (Primitive Shapes) Low-Level Parameters (Color Table) Application christof rezk-salama, institute for vision and graphics, university of siegen

Semantic Models Restrict ourselves to one specific application scenario. Example: CT angiography from neuroradiology The visualization task will be performed manually for multiple data sets. Visualization expert and medical doctor! Evaluate statistical information about the results: Which parameter modifications are necessary to „make the blood vessels sharper? “ Use dimensionality reduction (PCA) to create a semantic model christof rezk-salama, institute for vision and graphics, university of siegen

Developing a Semantic Model Step 1: Create a template for the TF Bone Brain/Soft Tissue Skin/Cavities Blood vessels christof rezk-salama, institute for vision and graphics, university of siegen

Developing a Semantic Model Step 2: Adapt the template to reference data christof rezk-salama, institute for vision and graphics, university of siegen

Developing a Semantic Model Step 2: Adapt the template to reference data christof rezk-salama, institute for vision and graphics, university of siegen

Developing a Semantic Model Step 3: Dimensionality Step 2: Adapt the template toreduction reference data Principal Component Analysis Semantics Reference Transfer Functions Semantic Model christof rezk-salama, institute for vision and graphics, university of siegen

Semantic Model High-Level User Interface High-Level Control Transfer Function Semantic Model christof rezk-salama, institute for vision and graphics, university of siegen

Semantic Model christof rezk-salama, institute for vision and graphics, university of siegen

Prototype Implementation Applicable to „anything that can be described by a parameter vector“ Take care of the scale! PCA for entire parameter vector is not appropriate Small details might be missed Our solution: • Split transfer function into entities (=structures, groups of primitives with same scale) • Perform PCA separately for each entity • Reassemble the transfer function christof rezk-salama, institute vision and graphics, university of siegen from thefordifferent entities

Results CTA: intracranial aneurysms: 512 x {120 -160} @12 bit, 100 ml non-ionic contrast dye 20 data sets for training / 5 data sets for evaluation MR brain surgery: 256 x {150 -200} @12 bit (noisy, lower dynamic range ~10 bit) 10 data sets Evaluation of the model: Analytically: Stability of the eigenvectors (dot product > 0. 9) Stable for >12 data sets (regardless of individual choice) User Study: Labels removed from the user interface Most semantics were correctly identified by non-expert users christof rezk-salama, institute for vision and graphics, university of siegen

Conclusion User Interface Design Strategies: Reducing DOF is not enough. Good user interfaces must provide an appropriate mental model Not an attempt to create a single user interfaces for any visualization tasks Create semantic models for examination tasks as specific as necessary Building block for software assistants for medical diagnosis and therapy planning christof rezk-salama, institute for vision and graphics, university of siegen

Acknowledgements Bernd Tomandl MD, Neuroradiologie, Bremen Christopher Nimsky MD, Neurochirurgie, Erlangen christof rezk-salama, institute for vision and graphics, university of siegen