1. Evaluation of a Bricked Volume Layout for a Medical Workstation based on Java Peter Kohlmann, Stefan Bruckner, Armin Kanitsar, M. Eduard Gröller Institute of Computer Graphics and Algorithms Vienna University of Technology

2. Peter Kohlmann 1 Outline Motivation Multi-planar Reformatting (MPR) Results for different access patterns MPR Random access Spherical access Conclusions

3. Peter Kohlmann 2 Motivation Most medical workstations: linear volume layout Increasing size of medical volume data Main memory: limiting factor for data visualization Better memory utilization with subdivided volumes Evaluation for company partner: use of bricked volume layout for medical workstation implemented in Java performance for common access patterns to medical volume data

4. Peter Kohlmann 3 Bricking in a Nutshell Medical volumes data sets: stacks of 2D images (slices) Linear volume layout: data values stored in single array Problem: rendering of large data sets Bricked volume layout: subdivision of volume into smaller parts (bricks) Single brick: fixed number of data values in x-, y- and z-dimension

5. Peter Kohlmann 4 Multi-Planar Reformatting in a Nutshell Important access pattern to medical volume data Radiologists prefer to examine 2D slices Arbitrary reformation of 2D image stack Medical workstations display volume data in different views

6. Peter Kohlmann 5 Basic Algorithms Brick Generation Image Brick Rasterization Basic Ray Setup Brick Prefetching Brick-wise Processing Ray Propagation MPR Computation

7. Peter Kohlmann 6 Brick Generation Efficient addressing: brick size power of two Good choice: 64 KB (32x32x32 x 16 bit) (Grimm et al. 04, Law and Yagel 96) Brick is simple data structure: unique ID min- and max-value padding

8. Peter Kohlmann 7 MPR Computation Brick-wise resampling of the volume along scan lines (rays) Ray Propagation Brick-wise Processing Brick Prefetching Basic Ray Setup Brick Rasterization

9. Peter Kohlmann 8 MPR Computation Ray Propagation Brick-wise Processing Brick Prefetching Basic Ray Setup Brick Rasterization Processing of a single brick Image generation

10. Peter Kohlmann 9 Results MPR Computation Random access Spherical access PC configuration AMD Athlon 64 Dual Core Processor 4400+ 2 GB of main memory NVIDIA GeForce 7800 GTX with 256 MB of internal memory Test data set: 512 x 512 x 333

11. Peter Kohlmann 10 Results – MPR Computation axial coronal sagittal arbitrary Computation time for single slice (512 x 512)

12. Peter Kohlmann 11 Results – MPR Computation Computation time for single slice (512 x 512) Evaluation: Axial and coronal: -30% Sagittal: +30% Randomly oriented plane: -16% High performance gaps for linear volume layout: different memory access patterns Utilization of bricking better data locality

13. Peter Kohlmann 12 Results – Random Access Worst case scenario to access data values concerning data locality Time to access 512 x 512 randomly distributed values 21.4 ms (linear volume layout) 41.4 ms (bricked volume layout) Calculation effort to access the data value at certain position Linear volume layout: one-level calculation Bricked volume layout: two-level calculation

14. Peter Kohlmann 13 Results – Spherical Access Definition of parameterized sphere inside volume Simulation of region growing Access 512 x 512 data values on parameterized sphere surface Radius: 5 to 150 Linear volume layout: 10.5 ms – 13.6 ms Bricked volume layout: No brick prefetching: 32 ms – 260 ms Brick prefetching: 15.5 ms (constant) reasonable performance with intelligent prefetching

15. Peter Kohlmann 14 Conclusions Evaluation results for different access patterns to medical volume data MPR Random access Spherical access Benefits of bricked volume layout more pronounced for larger data sets We recommend the application of bricked volume layout to a medical workstation based on Java

16. Peter Kohlmann 15 Thanks for your Attention!