Decades of research has led to the development of accurate, reliable, and fully automated image registration algorithms, the most recognized of which is intensity-based registration by maximizing mutual information (MI) between a pair of images. Although applicable to a range of modalities and organs (rigid or nonrigid) and therefore quite versatile, the slow execution of MI-based registration continues to limit its full clinical potential.

We present here a hardware-accelerated implementation of MI-based image registration that has reduced the time of registration from hours to less than a minute. Our implementation executes a well-known and well-validated nonrigid image registration algorithm[1] on three field-programmable gate array (FPGA) chips, resident on a single, commercially available add-on PC board. Because rigid registration is the first step in nonrigid registration, our implementation is able to compute rigid registration as well. Highly optimized pipelined and parallel computing along with a custom parallel memory architecture that significantly reduces data transfer overhead have helped achieve over 100-fold acceleration. The implementation has been optimized to produce results comparable to those from an equivalent software-only implementation.[2]

For characterization of speed, images were resized to a 256 x 256 x 256 matrix. Often this meant upsampling in the slice direction and cropping and/or downsampling slices to have 256 x 256 within-slice voxels . This resampling did not affect the registration accuracy significantly, because the registration accuracy is decided, to a large degree, by the largest voxel dimension in the original data. Our achievement of subvoxel accuracy reported below confirms this. The time of image registration was < 1 min (see Table 1).

Table 1. Time of image registration

We assessed registration accuracy by comparing our results with the reference results. For the Vanderbilt brain images, the reference results were ,in fact, the ground truth determined from known fiducial markers (although erased from the images). For the two nonrigid image registration databases, no ground truth existed. For these images, we treated previously validated, albeit slower, single-FPGA implementation as providing the reference solution. No significant difference was seen in the results of the two implementations (see Table 2). In each case, we normalized the observed difference by the largest voxel dimension (often the slice spacing) of the images in the original resolution prior to any resizing. This presented a unified approach to assess accuracy in spite of variations in voxel dimensions.

Table 2. Accuracy of image registration

Hardware-accelerated execution did not compromise the quality of image registration. A normalized registration accuracy of < 1 represents achievement of subvoxel accuracy (i.e., the images are registered with an error smaller than the largest voxel dimension). A value of 1, conversely, represents voxel-order accuracy. Subvoxel accuracy images was achieved for the first two databases. For the exhale-inhale CT database, the normalized registration accuracy was voxel-order. Because most registration algorithms aim to achieve subvoxel or voxel-order accuracies, our results indicate that we retained registration accuracy with the hardware implementation.

Finally, the use of images from various different modalities, and of various different organs, demonstrates the general-purpose nature of our image registration technology. We have incorporated the developed technology as a plug-in into OrisiX, a well-known medical imaging visualization software. OrisiX thus serves as the front-end of our technology. Our successful OsiriX integration indicates that the described technology can be integrated into most picture archiving and communication systems (PACS) to give radiologists easy access to the high-speed image registration capability described here.

[2] Dandekar O, Shekhar R. FPGA-accelerated deformable registration for improved target-delineation during CT-guided interventions. IEEE Trans on Biomedical Circuits and Systems. 2007;1(2):116-127.

[3] Shekhar R, Walimbe V, Raja S, et al. Automated 3-dimensional elastic registration of whole-body PET and CT from separate or combined scanners. Journal of Nuclear Medicine. 2005;46(9):1488-1496.

[4] Shekhar R, Lei P, Castro-Pareja CR, Plisker WL, D’Souza WD. Automatic segmentation of phase-correlated CT scans through nonrigid image registration using geometrically regularized free-form deformation. Medical Physics. 2007;34(7):3054-3066.