Real-time acquisition of 3D volumes is an emerging trend in medical imaging. True real-time 3D ultrasonic imaging is particularly valuable for echocardiography and trauma imaging as well as an intraoperative imaging technique for surgical navigation. Since the frame rate of ultrasonic imaging is fundamentally limited by the speed of sound, many schemes of forming multiple receive beams with a single transmit event have been proposed. With the advent of parallel receive beamforming, several architectures to form multiple (4-8) scan lines at a time have been suggested. Most of these architectures employ uniform sampling and input memory banks to store the samples acquired from all the channels. Some recent developments like crossed electrode array, coded excitation, and synthetic aperture imaging facilitate forming an entire 2D plane with a single transmit event. These techniques are speeding up frame rate to eventually accomplish true real-time 3D ultrasonic imaging. We present an FPGA-based scalable architecture capable of forming a complete scan plane in the time it usually takes to form a single scan line. Our current implementation supports 32 input channels per FPGA and up to 128 dynamically focused beam outputs. The desired focusing delay resolution is achieved using a hybrid scheme, with a combination of nonuniform sampling of the analog channels and linear interpolation for nonsparse delays within a user-specified minimum sampling interval. Overall, our pipelined architecture is capable of processing the input RF data in an online fashion, thereby reducing the input storage requirements and potentially providing better image quality.
We present an elastic registration algorithm based on local deformations modeled using cubic B-splines and controlled using 3D ChainMail. Our algorithm eliminates the appearance of folding artifacts and allows local rigidity and compressibility control independent of the image similarity metric being used. 3D ChainMail propagates large internal deformations between neighboring B-Spline control points, thereby preserving the topology of the transformed image without requiring the addition of penalty terms based on rigidity of the transformation field to the equation used to maximize image similarity. A novel application to virtual colonoscopy is presented where the algorithm is used to significantly improve cross-localization between colon locations in prone and supine CT images.
Fast computation of 3D deformation fields is critical to bringing the application of automated elastic image registration algorithms to routine clinical practice. However, it lies beyond the computational power of current microprocessors; therefore requiring implementations using either massively parallel computers or application-specific hardware accelerators. The use of massively parallel computers in a clinical setting is not practical or cost-effective, therefore making the use of hardware accelerators necessary. We present a hardware pipeline that allows accelerating the computation of 3D deformation fields to speeds up to two orders of magnitude faster than software implementations on current workstations and about 64 times faster than other previously reported architectures. The pipeline implements a version of the free-form deformation calculation algorithm, which is optimized to minimize the number of arithmetic operations required to calculate the transformation of a given set of neighboring voxels, thereby achieving an efficient and compact implementation in hardware which allows its use as part of a larger system.
Image similarity-based image registration is an iterative process that, depending on the number of degrees of freedom in the underlying transformation, may require hundreds to tens of thousands of image similarity computations to converge on a solution. Computation time often limits the use of such algorithms in real-life applications. We have previously shown that hardware acceleration can significantly reduce the time required to register two images. However, the hardware architectures we presented were limited to mutual information calculation, which is one of several commonly used image similarity measures. In this article we show how our architecture can be adapted for the calculation of other image similarity measures in approximately the same time and using the same hardware resources as those for the mutual information case. As in the case of mutual information calculation, the joint histogram is calculated as a first step. The image similarity measures considered are mutual information, normalized mutual information, normalized cross correlation, mean-square sum of differences and ratio image uniformity. We show how all these image similarities can be calculated from the joint histogram in a small fraction of the time required to calculate the joint histogram itself.
Mutual information is currently one of the most widely used image similarity measures for multimodality image registration. An important step in the calculation of the mutual information of two images is the estimation of their joint histogram. Most algorithms use lateral joint histogram sizes that are smaller than the actual number of intensity levels present in the images being registered. Using a reduced joint histogram size is especially useful when registering small portions of the images to obtain local deformations in nonrigid registration algorithms, and when implementing hardware solutions for acceleration of mutual information calculation. The most commonly used method for reducing the size of the joint histogram is to perform a linear rescaling of intensity values. The main problem with this method is that image regions with similar intensity values but corresponding to distinct tissue types tend to merge, thus compromising the accuracy of registration. We present new algorithms for reducing the number of gray levels present in 3D medical images, and compare their performance with previously reported ones. The tested algorithms are classified in three categories: histogram shape preserving algorithms, entropy maximization algorithms and quantization error minimization algorithms. Results show that in CT and MRI registration the best accuracy is achieved using entropy maximization algorithms, while in PET and MRI registration the best accuracy is achieved using histogram shape preservation algorithms.
Median Filtering and Convolution operations constitute 60-70% of the preprocessing operations performed on digital images. Software implementations of 3D filters in general-purpose processors do not match the speed requirements for real-time performance. Field Programmable Gate Arrays (FPGAs) support reconfigurable architectures that are sufficiently flexible to implement more than one operation in the existing hardware, yielding higher speed for real-time execution. We present a linear systolic array architecture for median filtering, that implements bit-serial searching and majority voting. The unique arrangement of line delay units endows parallelism to the bit-serial median finding algorithm. Convolution operation, based on the fast embedded multiplier units in the FPGA and an optimized Carry Save Adder array is also presented. The application of the above designs to 3D image preprocessing is described. A voxel rate of 220MHz is achieved for median filtering and 277MHz for convolution operation.
Three-dimensional ultrasonic imaging, especially the emerging real-time version of it, is particularly valuable in medical applications such as echocardiography and surgical navigation. A known problem with ultrasound images is their high level of speckle noise. Anisotropic diffusion filtering has been shown to be effective in enhancing the visual quality of 3D ultrasound images and as preprocessing prior to advanced image processing. However, due to its arithmetic complexity and the sheer size of 3D ultrasound images, it is not possible to perform online, real-time anisotropic diffusion filtering using standard software implementations. We present an FPGA-based architecture that allows performing anisotropic diffusion filtering of 3D images at acquisition rates, thus enabling the use of this filtering technique in real-time applications, such as visualization, registration and volume rendering.
Real-time image registration is potentially an enabling technology for the effective and efficient use of many image-guided diagnostic and treatment procedures relying on multimodality image fusion or serial image comparison. Mutual information is currently the best known image similarity measure for multimodality image registration. Mutual information calculation is a memory-intensive task that does not benefit from cache-based memory architecture in standard software implementations (i.e., the calculation incurs a large number of cache misses). Previous attempts to perform image registration in real time focused on parallel supercomputer implementations, which achieved real-time performance using large, expensive supercomputers that are simply impractical for deployment in a hospital. We have developed a custom hardware architecture that, in a single-module PC-based implementation, achieves registration speeds comparable to those of a 64-processor parallel supercomputer. The single-module speedup results from using parallel memory access and parallel calculation pipelines. The total speedup can be increased by using several modules in parallel. The architecture is designed for linear, mutual information-based registration and can be extended to elastic (nonlinear) registration.