GPU BASED IMPLEMENTATION OF SENSE (A PARALLEL MRI ALGORITHM) USING LEFT INVERSE METHOD

Abstract

A method including: constructing coil sensitivity encoding matrix; inversing of the coil sensitivity encoding matrix using Left Inverse method; and multiplying inverse of coil sensitivity encoding matrix with an under-sampled data matrix using a GPU residing on a host computer.

Claims

1. A method, comprising: constructing a coil sensitivity encoding matrix; inversing of the coil sensitivity encoding matrix using a Left Inverse method; and multiplying an inverse of the coil sensitivity encoding matrix with an under-sampled data matrix using a Graphical Processing Unit (GPU) residing on a host computer.

2. The method of claim 1 which is implemented on GPU to exploit maximum parallelism using a parallel approach.

3. The method of claim 2, further comprising computing all independent tasks on GPU by utilizing a maximum number of threads.

4. The method of claim 1, further comprising acquiring the under-sampled data by skipping k-space lines.

5. The method of claim 1 further comprising reconstructing Magnetic Resonance (MR) images by performing the inversion of coil sensitivity information.

6. The method of claim 1, further comprising reconstructing Magnetic Resonance (MR) images from the under-sampled data acquired from MM scanner having multiple receiver coils.

7. The method of claim 6, wherein an acceleration factor is less than the number of the multiple receiver coils.

8. The method in claim 1, wherein MR signals are used and are acquired by Cartesian sampling.

9. A system, comprising: a Magnetic Resonance Image (MM) scanner and the host computer comprising the GPU, wherein the data acquired from the MRI scanner is processed by the GPU by applying the method of claim 1.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings, which are included to provide a further understanding of the inventive concept, and are incorporated in and constitute a part of this specification, illustrate exemplary embodiments of the inventive concept, and, together with the description, serve to explain principles of the inventive concept.

[0027] FIG. 1A is an overview of CUDA Memory Model.

[0028] FIG. 1B is an example of CUDA grid organization.

[0029] FIG. 2 is a diagram to show the main steps of the research work.

[0030] FIG. 3 shows the tasks of one thread in CUDA by using left inverse method to invert a 8×2 matrix.

[0031] FIG. 4A is a fully sampled image.

[0032] FIG. 4B to FIG. 4E are SENSE reconstructions by CPU using 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils.

[0033] FIG. 4F to FIG. 4I are SENSE reconstructions by GPU using 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils.

[0034] FIG. 5A is a fully sampled image.

[0035] FIG. 5B to FIG. 5E are SENSE reconstructions by CPU using 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils.

[0036] FIG. 5F to FIG. 5I are SENSE reconstructions by GPU using 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils.

[0037] FIG. 6A is a SENSE Reconstruction time by CPU.

[0038] FIG. 6B is a SENSE Reconstruction time by GPU.

DETAILED DESCRIPTION OF ILLUSTRATED EMBODIMENTS

[0039] In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of various exemplary embodiments. It is apparent, however, that various exemplary embodiments may be practiced without these specific details or with one or more equivalent arrangements. In other instances, well-known structures and devices are shown in block diagram form in order to avoid unnecessarily obscuring various exemplary embodiments.

[0040] Unless otherwise specified, the illustrated exemplary embodiments are to be understood as providing exemplary features of varying detail of various exemplary embodiments. Therefore, unless otherwise specified, the features, components, modules, layers, films, panels, regions, and/or aspects of the various illustrations may be otherwise combined, separated, interchanged, and/or rearranged without departing from the disclosed exemplary embodiments. Further, in the accompanying figures, the size and relative sizes of layers, films, panels, regions, etc., may be exaggerated for clarity and descriptive purposes. When an exemplary embodiment may be implemented differently, a specific process order may be performed differently from the described order. For example, two consecutively described processes may be performed substantially at the same time or performed in an order opposite to the described order. Also, like reference numerals denote like elements.

[0041] For the purposes of this disclosure, “at least one of X, Y, and Z” and “at least one selected from the group consisting of X, Y, and Z” may be construed as X only, Y only, Z only, or any combination of two or more of X, Y, and Z, such as, for instance, XYZ, XYY, YZ, and ZZ. As used herein, the term “and/or” includes any and all combinations of one or more of the associated listed items.

[0042] In exemplary embodiments, modules and/or one or more components thereof, may be implemented via one or more general purpose and/or special purpose components, such as one or more discrete circuits, digital signal processing chips, integrated circuits, application specific integrated circuits, microprocessors, processors, programmable arrays, field programmable arrays, instruction set processors, and/or the like.

[0043] According to one or more exemplary embodiments, the features, functions, processes, etc., described herein may be implemented via software, hardware (e.g., general processor, digital signal processing (DSP) chip, an application specific integrated circuit (ASIC), field programmable gate arrays (FPGAs), etc.), firmware, or a combination thereof. In this manner, modules and/or one or more components thereof may include or otherwise be associated with one or more memories (not shown) including code (e.g., instructions) configured to cause modules, processors, and/or one or more components thereof to perform one or more of the features, functions, processes, etc., described herein.

[0044] The memories may be any medium that participates in providing code to the one or more software, hardware, and/or firmware components for execution. Such memories may be implemented in any suitable form, including, but not limited to, non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical or magnetic disks. Volatile media include dynamic memory. Transmission media include coaxial cables, copper wire and fiber optics. Transmission media can also take the form of acoustic, optical, or electromagnetic waves. Common forms of computer-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, any other magnetic medium, a compact disk-read only memory (CD-ROM), a rewriteable compact disk (CD-RW), a digital video disk (DVD), a rewriteable DVD (DVD-RW), any other optical medium, punch cards, paper tape, optical mark sheets, any other physical medium with patterns of holes or other optically recognizable indicia, a random-access memory (RAM), a programmable read only memory (PROM), and erasable programmable read only memory (EPROM), a FLASH-EPROM, any other memory chip or cartridge, a carrier wave, or any other medium from which information may be read by, for example, a controller/processor.

[0045] The idea behind SENSE is to reconstruct a un-aliased image from under-sampled data using the sensitivity encoding matrix obtained from the receiver coils. To implement SENSE, certain requirements need to be fulfilled: (1) coils sensitivity maps must be known, which may be obtained using Pre-scan method, (2) For accelerating the data acquisition in MRI, some phase encoding steps need to be skipped. The skipped phase encoding steps (k-space lines) decide the acceleration factor which affects the field of view (FOV).

[0046] The SENSE equation is given as:

M=[(C.sup.tψ.sup.−1C).sup.−1ψ.sup.−1]I  Equation 1

[0047] Where M is the image matrix to be reconstructed, C is the encoding matrix, ψ is the noise correlation matrix, I is the aliased image matrix. SENSE reconstruction algorithm requires inverting large number of independent small encoding matrices (C) which is time consuming if done sequentially. However, GPUs can be used to perform this task in parallel thus reducing the time for SENSE reconstruction. Also, MR signals are used and are acquired by Cartesian sampling.

[0048] FIG. 1A is an overview of CUDA Memory Model and FIG. 1B is an example of CUDA grid organization. Modern GPUs contain several multi-processors; each consists of many processor cores and high speed memory. GPU architecture supports Single Program Multiple Data (SIMD) programming model which allows programmers to easily implement the parallel algorithms on GPUs. The programmer can write single program which encompasses both the host (CPU) code and device (GPU) code. The NVIDIA compiler separates the host and device code which are then compiled by the host compiler and device compiler, respectively. The host code can transfer data to and from the GPU global memory. The GPU acts as a co-processor to the CPU and is also called General Purpose GPU (GPGPU). GPGPUs can perform calculations by launching a large number of threads where each thread processes a small part of the information in a parallel fashion. This allows splitting a big task into small portions and executing them in parallel by CUDA threads to reduce the execution time which is necessary in MRI. Threads are organized in thread blocks, and CUDA grid contains multiple thread blocks. Threads, blocks and grids each have unique identification number. CUDA Programmer decides how many grids, blocks and threads need to be initialized which depends on the underlying problem to be solved. All the threads start by reading from the device global memory, perform the required calculation and write back the results in the global memory. Registers, shared memory, L1 cache, L2 cache, constant memory support the programmer to write an efficient code for GPGPU. An overview of CUDA memory model and CUDA grid organization is shown in FIG. 1.

[0049] FIG. 2 is a diagram to show the main steps of the research work. According to an exemplary embodiment, SENSE algorithm is implemented in CUDA for GPU implementation. The software platform is MSDN 2010 and NVIDIA CUDA toolkit v6.5 is integrated in MSDN 2010. CUDA code is executed on GPU model NVIDIA GeForce GTX 660 with 960 CUDA cores, 2 GB memory with GDDR5 interface and 1033 MHz clock speed. The performance of SENSE algorithm on GPU is compared with its C language implementation for CPU core I7 with 2.9 GHz clock speed and 4 GB RAM. The work flow of this research is shown in FIG. 2.

[0050] The CPU implementation is very similar to the GPU implementation. The number of operations is exactly the same in both cases. The CPU implementation executes the code sequentially by using ‘For’ loops whereas GPU implementation breaks up the tasks and launches the threads to execute the tasks in parallel.

[0051] FIG. 3 shows the tasks of one thread in CUDA by using left inverse method to invert 8×2 matrix. In the GPU, a number of threads are launched as per matrix inversions required for SENSE algorithm to attain the highest level of parallelism in GPUs. Each thread is capable of selecting pixel values from the coil sensitivity matrices and the under-sampled image, and performs the matrix inversion. Moreover, it performs other required multiplications as per requirement of the left inverse method and computes the values for the reconstructed image. FIG. 3 elaborates the working of a CUDA thread and the details of the left inverse method as used in this paper. The temporary variables for thread usage are stored in registers instead of global memory to reduce the reconstruction time. The total time of execution of SENSE algorithm including writing the output in GPU global memory is measured. There are different ways to measure the computation time in CUDA which includes CPU timers and CUDA events however CUDA events are more reliable than CPU timers. CUDA events are used in this work to calculate the computation time of SENSE algorithm in GPU. In C language, the execution of SENSE algorithm is measured in milliseconds by using time functions.

[0052] To monitor the reconstructed image quality for both the CPU and GPU implementations, we have used artefact power as a quantifying parameter. Mean SNR is calculated using pseudo multiple replica method.

[0053] The experiments are performed on two datasets acquired from 1.5 Tesla MRI Scanner which includes phantom and human head data. The dimensions of the aliased images (under-sampled by factor 2) are 128×256×2, 128×256×4, 128×256×6, and 128×256×8 for two, four, six and eight receiver coils respectively.

[0054] FIG. 4A is a fully sampled image, FIG. 4B to FIG. 4E are SENSE reconstructions (using AF=2) by CPU using the data acquired from 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 coils, FIG. 4F to FIG. 4I are SENSE reconstructions (using AF=2) by GPU using the data acquired from 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils, FIG. 5A is a fully sampled image, FIG. 5B to FIG. 5E are SENSE reconstructions by CPU using the data acquired from 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils, FIG. 5F to FIG. 5I are SENSE reconstructions by GPU using the data acquired from 2 receiver coils, 4 receiver coils, 6 receiver coils and 8 receiver coils. FIGS. 4 and 5 show the SENSE reconstruction results obtained from both CPU and GPU implementations for phantom and human head data respectively. A summary of Artefact power of the CPU and GPU reconstructed images of both Phantom and human head data sets is shown in Table 1.

[0055] FIG. 6A is a SENSE Reconstruction time by CPU, FIG. 6B is a SENSE Reconstruction time by GPU. In GPU implementation, most of the time is consumed in transferring the data from the host (CPU) memory to the device (GPU) memory. Therefore, two time measurements are performed as shown in Table 2: (1) Time taken exclusively for SENSE reconstruction by the kernel, (2) Time taken by the kernel and data transfer time (transferring input data from the host memory to the device memory and copying back the results from the device memory to the host memory). The results from Table 2 show that GPU attains approximately 18ט28× performance improvements in terms of time consumption by kernel for SENSE reconstruction algorithm exclusively, and approximately 7ט10× performance improvements for kernel as well as data transfer; this clearly demonstrates the efficiency of GPU over CPU because of parallel executions. FIG. 6 and Table 2 compare the time consumption by GPU and CPU for different number of receiver coils used for SENSE reconstruction.

[0056] Mean SNR is found using the SNR map method and is shown in Table 3 for phantom image and human head images respectively. The mean SNR in phantom reconstructed image is more than 39 dB and for human head data it is more than 29 dB which shows that the reconstructed MR images possess good SNR.

[0057] The results according to one exemplary embodiment show that GPU implementation of SENSE has reduced the computation time significantly as compared to CPU implementation as well as the quality of the reconstructed image is maintained. The results show that the computation time increases with the number of receiver coils because the required number of operations (multiplications and additions) is increased.

[0058] The exemplary embodiments present the implementation of SENSE algorithm on GPU using left inverse method. The performance comparison of GPU and the multi core CPU implementations is performed. The rectangular matrix inversion is implemented in CUDA for GPU implementation of SENSE. The results according to one exemplary embodiment show that GPU provides approximately 7ט28× reduction in computational time as compared to the CPU. The future work includes the performance comparison between CPU and GPU for higher acceleration factors in SENSE. Also, with new generations of graphic cards, it will be possible to further reduce the computation time with better optimized GPU programs.

TABLE-US-00001 TABLE 1 Summary of AP for phantom image and human head image for CPU and GPU implementation of SENSE algorithm from FIG. 4 and 5. AP for AP for AP for No. of Phantom Phantom AP for Human Human receiver image using image using Head image using Head image coils multi-core CPU GPU multi-core CPU using GPU 2 3.07 × 10.sup.−31 2.45 × 10.sup.−13 3.07 × 10.sup.−31 5.94 × 10.sup.−13 4 2.69 × 10.sup.−31 2.60 × 10.sup.−13 2.70 × 10.sup.−31 5.99 × 10.sup.−13 6 2.34 × 10.sup.−31 2.54 × 10.sup.−13 2.34 × 10.sup.−31 6.06 × 10.sup.−13 8 9.35 × 10.sup.−32 2.53 × 10.sup.−13 9.35 × 10.sup.−32 5.99 × 10.sup.−13

TABLE-US-00002 TABLE 2 Performance Comparison of multi core CPU and GPU. No. of SENSE SENSE receiver coils SENSE computation time computation (Acceleration Computation time by GPU (kernel time by GPU factor = 2) by multi-core CPU and data transfer) (kernel only) 2 14 ms 1.3 ms 0.5 ms 4 18 ms 2.6 ms 1.0 ms 6 31 ms 3.1 ms 1.4 ms 8 47 ms 4.7 ms 1.7 ms

TABLE-US-00003 TABLE 3 Mean SNR for phantom image and human head image for CPU and GPU implementation of SENSE algorithm Mean SNR for Mean Mean SNR for Mean SNR Human Head SNR for No. of Phantom image for Phantom image Human Head receiver using multi-core image using using multi- image using coils CPU (dB) GPU (dB) core CPU (dB) GPU (dB) 2 39.4721 39.4699 29.1428 29.2865 4 39.93 39.9256 28.9766 29.210 6 39.4723 39.4680 29.294 29.2858 8 39.4688 39.4742 29.2869 29.2876

[0059] Although certain exemplary embodiments and implementations have been described herein, other embodiments and modifications will be apparent from this description. Accordingly, the inventive concept is not limited to such embodiments, but rather to the broader scope of the presented claims and various obvious modifications and equivalent arrangements.