Title: Full Body Virtual Autopsies using a State-of-the-art Volume Rendering Pipeline
1Full Body Virtual Autopsiesusing a
State-of-the-artVolume Rendering Pipeline
- Patric Ljung NVIS/VITA, LiU
- Calle Winskog RMV
- Anders Persson CMIV, LiU
- Claes Lundström CMIV, LiU Sectra Imtec AB
- Anders Ynnerman NVIS/VITA, LiU
2Forensic Autopsy Workflow
CT Scan
Crime Scene Investigation
Police
Radiology
DVR Visualization
Physical Autopsy
Forensics
Visual Exploration
3Siemens Definition CT Scanner
Scan length 2 m 10 GB under a minute
4Challenge of Virtual Autopsies
- Gigabytes of data do not fit in memory
- Full body and high resolution details
- Interaction and run-time TF changes
5 GB data set (512x512x3396), Data reduction 301
(170 MB)LOD distribution 177709, 19, 15487,
20292, 4605 (total 218112 blocks)
5DVR Pipeline
Out-of-Core Data Management
TF
Large Data Sets
Significant Data Reduction
STORAGE
LOADER
ANALYSER
RENDERER
Interactive
High Image Quality
LOD
Dual TF Raycasting
Flat Volume Blocking
Interblock Interpolation
TF-based LOD
Adaptive Volume Sampling
6Multiresolution Volumes
- Hierarchical blocking
- Constant data size
- Lower resolution blocks cover increasing spatial
size
- Flat blocking structure
- Constant spatial size
- Lower resolution blocks have fewer samples
Level 0
Level 0
Level 1
Level 1
Level 2
Level 3
Level 2
Level 3
7Hierarchical vs Flat Blocking
Hierarchical blocking
Flat blocking
Final composition
Final composition
Data reduction 2.21
Data reduction 3.81
8TF Based Level-of-Detail
- Block resolution derived by TF content
- Content of a block in the current TF domain
- Empty All voxels are transparent
- Homogeneous Non-transparent voxels, similar
color - Varying Varying transparency, varying color
- Block significance based on visual error
- Color difference from CIE Luv, DE
- Perceptually uniform
- Optimize LOD to minimize visual error for a given
memory limit
9TF Based LOD Data Reduction
CIE Luv DE error
Data Reduction
6411.6 (2.25 MB)
20010.5 (0.72 MB)
40010.25 (0.36 MB)
6411.6 (2.25 MB)
8112.5 (18 MB)
144 MB, 512x512x384
10 Multiresolution Interblock Interpolation
Situation A blocked multiresolution
representation
Problem The domain between block sample
boundaries
Required Smooth interpolation of samples between
blocks
11Multiresolution Interblock Interpolation
Data reduction 801 (1.25 of original data size)
Without Interblock Interpolation
With Interblock Interpolation
12Adaptive Volume Sampling
13Adaptive Volume Sampling
1024x1024
Adaptive density ATI 14.1 FPS NV 6.0
FPS Speed-up 2.6-2.8
Full density ATI 5.1 FPS NV 2.3 FPS
- ATI X1800 XT 512MB - NV GF7800 GTX 256MB
14Dual TF Rendering
1024x1024
9.2 / 2.0 FPS
7.5 / 1.8 FPS
15Out-of-Core Data Management
- Disk performance
- Data transfer rates, about 60 MB/s
- Access density performance is still poor!!!
- Random access of many small blocks is bad
- Group of blocks GOBs
- Large pages 24 192 kb
- Spatial coherence
- Minimizes disk read requests
- Uncompressed storage
- Precomputed gradients
16Out-of-Core Data Management
17Man Killed by Police Officer
- Man acts aggressively when police arrives
- Police officer fires gun, man is dead on arrival
- Police officer accused for not warning
- Warning shot ricocheted as indicated by bullet
fragments
18Routine Procedure Gone Wrong
- Routine procedure to control respiration for
surgery - A needle was inserted to remove air from the lung
cavity (sack) - Accidentally the heart stopped
- Physical autopsy fails to explain the cause of
death - Virtual autopsy reveals air in unexpected parts
of the body
19Conclusions
- The Virtual Autopsy procedure
- Enables digital exploration of human cadavers
- Is already an important forensic tool
- May in some cases replace traditional autopsies
- Honors religious boundaries and demands from
relatives - Full body volume rendering on commodity PCs
- Graceful adaptation to available resources
- High quality LOD and interactivity at high
resolutions - Preserving full quality of volume data
20Thank You
5 GB data set, Data reduction 301 (170 MB),
Image resolution 1920x560
Efficient Methods for Direct Volume Rendering of
Large Data Sets PhD Thesis, P. Ljung, 2006
(Google Patric Ljung) http//www.itn.liu.se/plg/p
hdthesis/diss.pdf