Some Applications of GPU-Based Medical Imaging Baohua Wu Roadmap • Introduction • Medical imaging applications – Decompression – Registration • Conclusion Introduction to GPU-based Medical Imaging • • • • Visualization Segmentation Registration Codec Source: Gianluca Paladini, State of the Art in GPU-Accelerated Techniques for Medical Imaging, GTC09 Motivations • Challenges from medical imaging – Large volume of data (gigabytes to terabytes) – Processing time on CPU (minutes, hours or even days) • Limitations of some hardware – parallel computers – FPGA, dedicated devices • GPU’s emergence offers a solution Visualization of Medical Images • • • • • • • • Automatic carving 4D flow visualization Diffusion tractography Virtual endoscopy (ex. artery) Virtual unfolding (ex. colon) Tissue classification Virtual mirrors etc Image Segmentation • “Segmentation refers to the process of partitioning a digital image into multiple segments” – wikipedia.org Source: Gianluca Paladini, State of the Art in GPU-Accelerated Techniques for Medical Imaging, GTC09 Image Registration Source: http://www.siam.org/meetings/op08/Modersitzki.pdf GPU-Accelerated Registration • Adaptive Radiation Therapy • Real-time ultrasound / CT registration Application 1 GPU-based Decompression for Medical Imaging Applications Albert Wegener GPU Technology Conference 2009 Faster Imaging System Problems & Solutions • Serial coding with VLC (Variable Length Code) – Data are stored in packets that can be decoded in parallel • Small shared memory prevents storing one entire packet per thread – n symbols at a time • Conditionals lead to divergent warps – Replace conditionals with lookup tables Data-driven look-up table Application 2 Medical Image Registration with CUDA Richard Ansorge GTC 09 Method • Deformation model: – Affine – B-spline • Search strategy – Simplex – Gradient descent • Cost function: – correlation coefficient – mutual information 2D histogram of intensities of two images • Source: F. E. M. S. Matthias Tessmann, Christian Eisenacher and P. Hastreiter. Gpu accelerated normalized mutual information and b-spline transformation. In Eurographics Workshop on Visual Computing for Biomedicine (EG VCBM), pages 117–124, 2008. Application 3 Fast deformable registration on the gpu: A cuda implementation of demons P. Muyan-Ozcelik, J. Owens, J. Xia, and S. Samant IEEE Conference on Computational Sciences and Its Applications, 2008 Demons Algorithm Source: J.-P. Thirion, Image matching as a diffusion process: an analogy with Maxwell’s Demons, MIA 98 Demons Algorithm • v: the displacement where S: the static image, M: the moving image, i: a position in the image • Similarity measure of Correlation Coefficient: where D: the deformed moving image Control flow graph of Demons algorithm • Source: X. Gu, H. Pan, Y. Liang, R. Castillo, D. Yang, D. Choi, E. Castillo, A. Majumdar, T. Guerrero, and S. B. Jiang. Implementation and evaluation of various demons deformable image registration algorithms on a gpu. Physics in Medicine and Biology, 55(1):207-219, 2010. CUDA Kernels Speedups Conclusion • GPU opens the prelude of a new era for medical imaging – Post-processing to real-time processing with speedups from tens to hundreds of times – More automated workflow in surgical operations – Interventional medical imaging – Adaptive radiation therapies Acknowledgement • • • • • Joseph T Kider Jr. Jonathan McCaffrey Gang Song Dr. Brian Avants Dr. James Gee