6
Digitally reconstructed
radiograph (DRR)
Coordinates mapping by simulating X-ray projection in
data space [Lemieux94]
SimulationData
space
Real
space
X-ray
source
X-ray image
Position of the
real patient can
be estimated if
the DRR
matches the
X-ray image
Projection
Preoperative
CT volume
2-D/3-D 位置合わせ
7.
GPU implementation7
GPU SideCPUSide
CT
volume
2-D
image
DRR
1) DRR
generation
Gradient
DRR
Gradient
image
2) Gradient image generation
Reduction
3) NCC computation
Similarity
evaluation
5 floating point numbers
P:
Position &
orientation
Input data
P
Minimize
communication
between CPU and GPU
Bottleneck
offloaded to GPU
Update
P
Data transfer
8.
DRR generation
Graphics-orientedsolution: DRR generation from CT
volume is almost the same computation as volume
rendering
We select the texture-based rendering method
[Cullip93], which exploits the hardware components in
the GPU
8
Viewing direction
CT volume
Volume slicing by
texture mapping
Screen
Slice composition by
alpha blending
9.
Performance results
GPU-basedsystem
Laptop PC (Dell Precision M70)
CPU: Pentium 4 2.8-GHz
GPU: NVIDIA Quadro FX Go 1400
Video memory: 256MB
Cluster-based system
Cluster of 64 PCs
CPU: Dual Pentium 3 1-GHz
Network: Myrinet-2000 (2 Gb/s)
9
1
GPU
1
CPU
128
CPUs
1) DRR (ms) 19 2940 142
2) Gradient
(ms)
4 142 7
3) NCC (ms) 28 9 46
Total per
iteration (ms)
51 3091 195
300 iterations 15 s 15 m 58 s
10.
Cone beam reconstruction
A technique for producing volume data
Input data: a series of 2-D projections taken by CT scanner
Output data: 3-D volume
Feldkamp, Davis and Kress (FDK) algorithm
Standard but memory-intensive reconstruction method
135 s on a single-core CPU using SSE instructions
Acceleration is needed for real-time imaging
GPU (8.9 s), Cell B.E. (19 s), and FPGA (40 s)
10
Imaging Reconstruction
A series of 2-D projections Volume
Scan rotation
Cone beam CT
11.
FDK algorithm
Input/outputdata
Input: a series of projections
Output: Volume
Filtering stage
Convolution between projections
and Shepp-Logan filter
Backprojection stage
accumulates pixels on filtered
projections to voxels
iterates accumulation for every projection
has perfect parallelism between different voxels
11
Volume F
X-ray source
Filtered
projection
Qi
u
v
Interpolated
pixel (u,v)
Rotation
F(x,y,z) = ∑ Wi(x,y) Qi(u,v)
Volume Weight Projectioni