Real-Time 2D-3D Medical Image Registration using RapidMind Multi-Core Development Platform

Proceedings of the 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC08), Vancouver, August 20-24, 2008.
(Co-author: Lin Xu)

In this research, we develop an efficient intensity-based rigid 2D-3D image registration method. We implement the resulting algorithm using the RapidMind Multi-core Development Platform to exploit the highly parallel multi-core architecture of graphics processing units (GPUs). We use a ray casting algorithm to generate the digitally reconstructed radiographs (DRRs) on GPUs and efficiently reduce the complexity of DRR construction. The registration optimization problem is solved by the Gauss-Newton method. To fully exploit the multi-core parallelism, we implement almost the entire registration process in parallel by RapidMind. Numerical results are presented to demonstrate the efficiency of our method.

Example 1: (artificial data)

The 3D image volume is a white cube. The template 2D image is simulated by projecting the 3D image volume with known transformation parameters.
The GPU is NVIDIA GeForce 8800GTX.

Image resolution C++ RapidMind
16x16x16 Time per iteration (sec) 0.015 0.138
Iteration 6 6
32x32x32 Time per iteration (sec) 0.200 0.186
Iteration 5 5
64x64x64 Time per iteration (sec) 5.308 0.262
Iteration 5 5
128x128x128 Time per iteration (sec) 56.720 0.400
Iteration 6 6

Example 2: (clinical data)

The 3D image volume is a CT scan of a tripod fracture of a skull.

Portal image parameters C++ RapidMind
Rotations: (2, 2, 2)
Translations: (2mm, 2mm, 2mm) 		Total time (sec) 168.16 1.75
Iteration 5 5
Rotations: (4, 4, 4)
Translations: (4mm, 4mm, 4mm) 		Time per iteration (sec) 263.30 2.70
Iteration 8 8
Rotations: (6, 6, 6)
Translations: (6mm, 6mm, 6mm) 		Time per iteration (sec) 371.39 3.65
Iteration 11 11