Acceleration of Medical Image Registration using Graphics Process Units in Computing Normalized Mutual Information

1. Acceleration of Medical Image Registration using Graphics Process Unitsin Computing Normalized Mutual Information Wei-Hung Cheng, Cheng-Chang Lu Department of Computer Science Kent State University

2. Introduction The computing capacities of graphics processing units (GPUs) have improved exponentially in the recent decade. NVIDIA released a CUDA programming model for GPUs. The CUDA programming environment applies the parallel processing capabilities of the GPUs to medical image processing research. Kent State University

3. Multi-resolution Approach Registration Why Multi-resolution? • Methods for detecting optimality can not guarantee that a global optimal value will be found. • Time to evaluate the registration criterion is proportional to the number of voxels. The result at coarser level is used as the starting point for the finer level. Using multi-resolution in conjunction with maximization of mutual information has been proven very helpful in registration processes. Multi-resolution is supplemented with binarization giving improved accuracy . Kent State University

4. A Binarization Registration Approaches • Background segmentation. • Linear binning size. • Using sub-sampling multi-resolution. • For CT-MR registration. Kent State University

5. Two stages: • Region Growing: Segmentation into Background and Foreground. 2. Two levels Registration: Binarized 2-bin images are input to the lower level. • Down-sampled binarized images as the input to the first level. • Result of the first level as the initial estimate for the second level. • The second level performs the registration of full images, using Maximization of Normalized Mutual Information. Kent State University

6. Normalized Mutual Information Mutual information It is applied to measure the statistic dependence between the image intensities of corresponding voxels in both images, which is assumed to be maximal if the images are geometrically aligned. Kent State University

7. Normalized Mutual Information Studholme et. al.: Where and • Compensate for the sensitivity of MI to changes in image overlap Kent State University

8. A Binarization ApproachRegion Growing Implementation • Finding Starting Points: Easy to select a point as seed for background. • Similarity Criteria: Threshold T can be extracted from the histogram Typical histogram for CT image (left) & MR image (right) Kent State University

9. A Binarization Approach • Median and maximum error between the prospective gold-standard and several retrospective registration techniques. Ours is labeled as LO1. • Median Error • Maximum Error Kent State University

10. Non-Linear Binning Approach • Background segmentation • Segmented image as input to K-means Clustering • Initially partition the image voxels into k bins where k = 256. 1a. Put all the background voxels into bin 0. 1b. Calculate the step size for the other k-1 bins using Each bin will be assigned all voxels whose intensity falls within the range of its boundary. 1c. Calculate the centroid of each bin. Kent State University

11. Non-Linear Binning Approach • K-means Clustering (cont.) • For each voxel in the image, compute the distances to the centroids of its current,previous, and next bin, if exists; if it is not currently in the bin with the closest centroid, switch it to that bin, and update the centroids of both bins. • Repeat step 2 until convergence is achieved; that is, continue until a pass through all the voxels in the image causes no new assignments or until a maximization iterations is reached where the maximization iterations = 500. • Two-level Registration Kent State University

12. Non-Linear Binning Approach • Median and maximum error between the prospective gold-standard and several retrospective registration techniques. Ours is labeled as LO2. • Median Error • Maximum Error Kent State University

13. Wavelet-based Multi-resolution Approach • Description: • Multi-resolution: Improve optimization speed and capture range. • The wavelet intends to transform images into a multi-scale representation. • A wavelet can be created by passing the image through a series of filter bank stages. Kent State University

14. Wavelet-based Multi-resolution Approach • Implementation: • Daubechies Wavelet filter coefficients ( DAUB4 ) • Four-level WT on 41 CT-MR pairs registration. • Three-level WT on 35 PET-MR pairs registration Kent State University

15. Wavelet-based Multi-resolution Approach • Result (cont.) A typical superposition of PET-MR images. Left : before registration Right: after registration. Kent State University

16. Wavelet-based Multi-resolution Approach • Median and maximum error between the prospective gold-standard and several retrospective registration techniques. Ours is labeled as LO3. • Median Error • Maximum Error Kent State University

17. Wavelet-based Multi-resolution Approach • Median and maximum error between the prospective gold-standard and several retrospective registration techniques. Ours is labeled as LO3. • Median Error • Maximum Error Kent State University

18. CUDA Programming Model • A parallel programming model and software environment designed to handle parallel computing tasks. • Similar to the traditional single instruction, multiple data (SIMD) parallel model. • Major abstractions: – a hierarchy of thread groups – shared memories – barrier synchronization • It provide a programming model for data parallelism, thread parallelism, and task parallelism. Kent State University

19. Computation paradigm of the CUDA A program is divided into blocks. – A block is a group of threads mapped to a single multiprocessor by the programmer to share the memory. The data is also divided amongst all threads in a SIMD fashion by the programmer. All threads are organized into warps. – Each warp is a group of 32 parallel scale threads, which can run concurrently on the multi-processors. Collections of warps are known as thread block Kent State University

20. Memory hierarchy is in the form of registers, constant memory, global memory, and textures. – registers: fastest level in the hierarchy, a limited amount of space. – constant memory: a subset of device memory, cannot be modified at run-time by a device. – global memory: permits read and write operation from all threads, but is uncached and has long latencies. – textures memory: a subset of the device memory, read-only on the device, faster cached reads, allows addressing through a specialized texture unit. Kent State University

21. CUDA Algorithm for NMI registration • The registration procedure spent 90-95% of its run-time on mutual information computation. • Medical image registration has a high level of data parallelism and image data can be mapped onto the GPUs platform. • Based on our normalized mutual information method, the tasks are classified into four CUDA kernels as follows: Transformation – This group performs coordinate transform, affine transform, and mapping matrix to establish spatial correspondence between two images. Interpolation – This group involves iteratively transforming image A with respect to image B while optimizing the MI measure which is calculated from corresponding voxel values. Histogram – This group computes a joint histogram of the pairs of images to evaluate the mutual information. Optimization – This group detects optimization of estimate transformation to evaluate its similarity Kent State University

22. The CUDA implementation consists of four stages: 1) Allocate data memory on the device and transfer them from the host to the device. 2) Set up the function kernel configuration. 3) Launch function kernel(s) and store the result in the device memory. 4) Transfer data from the device memory to the host memory. Kent State University

23. CUDA Implementation Experimental Results • The experiments involved the data sets of 7 patients, each consisting of Computed Tomography (CT) and six Magnetic-Resonance (MR) volumes. • On a PC, having a 2.40 GHz Intel® Core™ 2 Quad CPUs, and 4 GB DDR2 memory with NVIDIA’s GeForce 9600 GT graphic card. • All CT images were registered to the MR images using the MR image as the reference image on PC • Run the registration procedure on both for the CPU-base platform (C program), and the GPUs platform (the CUDA program). • Experimental results showed that the GPU implementation improves the registration computational performance with a speedup factor of 23.4× Kent State University

24. Table: Data Set Image Characteristics Kent State University

25. Run Time Range for 41 pairs data set on CPU and GPU Comparison of GPU and CPU-based implementation for the registration procedure runtimes Kent State University

26. References • A. Collignon, F. Maes, D. Delaere, D. Vandermeulen, P. Suetens, and G. Marchal. Automated multi-modality image registration based on information theory. Information Process. Med. Imaging, pages 263–274, 1995. • R. Gonzalez and R. Woods. Digital Image Procesing. Prentice Hall Press, 2 edition, 2002. • M. Harris. Mapping computational concepts to GPUs. GPU Gems 2, chapter 31. Addison Wesley, Mar. 2005. • F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, and P. Suetens. Multimodality image registration by maximization of mutual information. IEEE Trans. Med. Imaging., 16(2):187–198, 1997. • J. B. A. Maintz and M. A. Viergever. A survey of medical image registration. Medical Image Analysis., 2(1):1–36, 1998. • NVIDIA. Nvidia cuda compute unified device architecture. Programming Guide, Version 2.0. NVIDIA, 2008. • J. D. Owens, M. Houston, D. Luebke, S. Green, J. E. Stone, and J. C. Phillips. Cpu computing. Proceedings of the IEEE, 96(5):879–899, 2008. • D. L. Pham, C. Xu, and J. L. Prince. Current methods in medical images segementation. Annual Review of Biomedical Engineering, 2, 2000. • W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling. Numerical recipes in c: The art of scientific computing. 2nded. Cambridge, U.K.:Cambridge University Press, pages 394–455, 1999. • C. Studholme, D. L. G. Hill, and D. J. Hawkes. An overlap invariant entropy measure of 3d medical image alignment. Pattern Recognition., 32(1):71–86, 1999. • W. M. WellIII, P. Viola, H. Atsumi, S. Nakajima, and R. Kikinis. Multi-modal volume registration by maximization of mutual information. Medical Image Analysis., Kent State University