GPU BASED IMPLEMENTATION OF SENSE (A PARALLEL MRI ALGORITHM) USING QR DECOMPOSITION

Abstract

A method of SENSE reconstruction including: constructing a coil sensitivity encoding matrix; inversing of the coil sensitivity encoding matrix using a QR decomposition algorithm; and multiplying an inverse of the receiver coil sensitivity encoding matrix with an under-sampled data using a central processing unit (CPU) and using a GPU residing on a host computer to further decrease computation time.

Claims

1. A method, comprising: constructing a coil sensitivity encoding matrix; inversing of the coil sensitivity encoding matrix using a QR decomposition algorithm; and multiplying an inverse of the coil sensitivity encoding matrix with an under-sampled data using a central processing unit (CPU).

2. The method of claim 1, wherein the step of multiplying is further implemented on a graphics processing unit (GPU) to exploit maximum parallelism using a parallel approach.

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

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 receiver coil sensitivity encoding matrix.

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

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

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

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

10. A system, comprising: a Magnetic Resonance Imaging (MRI) scanner and a computer comprising the CPU and the GPU, wherein the data acquired from the MRI scanner is processed by the GPU residing on a host computer by applying the method of claim 2.

11. A method, comprising: constructing a coil sensitivity encoding matrix; inversing of the coil sensitivity encoding matrix using a QR decomposition algorithm; and multiplying an inverse of the coil sensitivity encoding matrix with an under-sampled data using a GPU.

12. The method of claim 11, further comprising computing all the independent tasks by utilizing a maximum number of kernels.

13. The method of claim 11, further comprising reconstructing Magnetic Resonance (MR) images by performing the inversion of the coil sensitivity encoding matrix.

14. The method of claim 11, further comprising acquiring the under-sampled data by skipping k-space lines.

15. The method of claim 11, further comprising reconstructing Magnetic Resonance (MR) data from the under-sampled data acquired from a Magnetic Resonance Image (MRI) scanner having multiple receiver coils.

16. The method of claim 15, wherein an acceleration factor is less than the number of the multiple receiver coils.

17. The method in claim 11, wherein MR signals are used and are acquired by Cartesian sampling.

18. A system, comprising: an MRI scanner and a computer comprising the GPU, wherein the data acquired from the MRI scanner is processed by the GPU by applying the method of claim 11.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0027] 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.

[0028] FIG. 1 is a pictorial representation of the main equation of SENSE.

[0029] FIG. 2 is a flowchart showing QR decomposition based SENSE reconstruction, first implemented on Central Processing Unit (CPU) using C-language, based on an exemplary embodiment;

[0030] FIG. 3 shows a flow chart of an implementation of QR decomposition based SENSE reconstruction on GPU using Compute Unified Device Architecture (CUDA) language;

[0031] FIG. 4A shows an original human head dataset image using QR based SENSE algorithm according to an exemplary embodiment, FIG. 4B shows the reconstructed image of human head dataset reconstructed using QR based SENSE algorithm on CPU according to an exemplary embodiment, and FIG. 4C shows the reconstructed image of human head dataset reconstructed using QR based SENSE algorithm on GPU according to an exemplary embodiment;

[0032] FIG. 5A shows an original image of simulated brain dataset using the method according to an exemplary embodiment, FIG. 5B shows the reconstructed image of simulated brain dataset using CPU according to an exemplary method, and FIG. 5C shows the reconstructed image of simulated brain dataset using GPU according to an exemplary method;

[0033] FIG. 6A shows the original image for comparison purpose of the phantom dataset images using QR based SENSE algorithm according to an exemplary embodiment; FIG. 6B shows the reconstructed phantom dataset image using QR based SENSE algorithm on CPU according to an exemplary embodiment, and FIG. 6C shows the reconstructed phantom dataset image using QR based SENSE algorithm on GPU according to an exemplary embodiment;

[0034] FIG. 7A shows the original image of cardiac dataset having 30 channel coils according to an exemplary embodiment, FIG. 7B shows the reconstructed image for undersampled 30 channel cardiac dataset using CPU according to an exemplary embodiment, and FIG. 7C shows the reconstructed image for undersampled 30 channel cardiac dataset using GPU according to an exemplary embodiment.

[0035] FIG. 8 shows a comparison of the time taken by CPU and GPU for the 8 receiver coil simulated human head and phantom datasets according to exemplary embodiments;

[0036] FIG. 9 gives a comparison of the time taken by CPU and GPU for the 30 receiver coil cardiac dataset according to an exemplary embodiment.

DETAILED DESCRIPTION OF THE ILLUSTRATED EMBODIMENTS

[0037] FIG. 1 is a pictorial representation of the main equation of SENSE:

{right arrow over (I)}=Ĉ*{right arrow over (ρ)}  (1)

[0038] Where {right arrow over (I)} represents the folded pixels; Ĉ is the receiver coil sensitivities and {right arrow over (ρ)} is the required image. The pictorial representation of this equation is shown in FIG. 1.

[0039] For equation 1, the equivalent mathematical matrix would be

[00001]$\left[\begin{array}{c}{\stackrel{->}{I}}_1\\ {\stackrel{->}{I}}_2\\ {\stackrel{->}{I}}_3\\ {\stackrel{->}{I}}_4\end{array}\right]=\left[\begin{array}{cccc}{C}_{11}& C_{12}& C_{13}& C_{14}\\ C_{21}& C_{22}& C_{23}& C_{24}\\ C_{31}& C_{32}& C_{33}& C_{34}\\ C_{41}& C_{42}& C_{43}& C_{44}\end{array}\right].Math.\left[\begin{array}{c}{\stackrel{->}{\rho }}_1\\ {\stackrel{->}{\rho }}_2\\ {\stackrel{->}{\rho }}_3\\ {\stackrel{->}{\rho }}_4\end{array}\right]$

[0040] In order to find the solution image, Ĉ (encoding matrix) must be inverted as given below:

{right arrow over (ρ)}={right arrow over (I)}Ĉ.sup.−1  (2)

[0041] It is to be noted that normally Ĉ matrix is not a square matrix, so its inverse cannot be calculated directly. Methods such as Cholesky factorization, QR decomposition, Left inverse method etc. can be used for the inversion of a rectangular matrix. In this work, QR decomposition based SENSE reconstruction is successfully implemented on CPU and GPU. The results show that SENSE reconstruction using GPU helps to significantly reduce the computational time. Of note, Q is an orthogonal matrix and R is an upper triangular matrix in linear algebra.

[0042] According to exemplary embodiments, QR decomposition is used to compute the inverse of the coil sensitivity encoding matrix ‘Ĉ’ in SENSE. Several methods exist to compute the QR decomposition. Due to the nature of the problem in this work Classical Gram-Schmidt projection based QR decomposition is used to implement SENSE reconstruction from the under sampled k-space data. QR algorithm decomposes a matrix C into matrices Q and R where C is an m×n rectangular matrix, Q is m×n orthogonal matrix and R is n×n upper triangular matrix.

[0043] FIG. 2 is a flowchart showing QR decomposition based SENSE reconstruction implemented on a Central Processing Unit (CPU) using C-language, based on an exemplary embodiment. In the first step matrix is decomposed into matrices Q and R. In the next step, inverse of the square matrix R (2×2) called R′ and Q.sup.t transpose of Q (8×2) is calculated. Afterwards, R′ is multiplied with Q.sup.t to get the resultant value for the two points. As code for the CPU is written in C-language that will process data sequentially. So, C-language code requires multiple iterations for complete image reconstruction. The number of iterations required to reconstruct image depends on the size of the image and acceleration factor.

[0044] FIG. 3 shows a flow chart of an implementation of the QR decomposition based SENSE reconstruction on GPU using Compute Unified Device Architecture (CUDA) language. Further reduction in image reconstruction time is achieved by such implementation. CUDA supports multiple thread execution in parallel. The loop in C language is easily unrolled for parallel implementation because there are no dependencies between iterations. One single thread is assigned against each C code iteration that will perform all the required steps for QR decomposition based SENSE reconstruction. Threads equal to the number of iterations in C implementation are launched at once to get maximum possible speedup and performance. However, the total number of operations performed in CUDA code for GPU is the same as C code for CPU. Due to this reason quality of the reconstruction results obtained from GPU is the same as in CPU but with almost 6× less time.

[0045] Exemplary embodiments present the implementation of QR based SENSE algorithm on CPU and GPU to reconstruct MR images from the acquired under-sampled data. The experiments are performed on different datasets. Simulated human head dataset having 8 channel head coils is used to test the feasibility of QR based SENSE approach.

[0046] The performance of QR based SENSE algorithm is evaluated using artifact power and reconstruction time. Artifact power is a measure of the square difference error between the reference image (I.sub.ref) and the reconstructed image (I.sub.recon). It is measured by Equation 2:

[00002]$\begin{array}{cc}\mathrm{AP}=\frac{\Sigma .Math.{.Math..Math.I_{\mathrm{ref}}.Math.-.Math.I_{\mathrm{recon}}.Math..Math.}^2}{\Sigma .Math.{.Math.I_{\mathrm{ref}}.Math.}^2}& \mathrm{Equation}.Math..Math.2\end{array}$

[0047] FIGS. 4A, 4B, and 4C are the human head dataset reconstructed images using QR based SENSE algorithm. FIG. 4A shows the original image, FIG. 4B shows the reconstructed image using CPU and FIG. 4C shows the reconstructed image using GPU. The time taken by CPU (Core 2 Duo 3.00 GHz) for human head data set is 200 ms because the inversion of Encoding matrix is performed iteratively. The implementation on GPU (Nvidia GTX 560 Ti) for human head data set takes 2.4 ms to reconstruct the same image as it performs the inversion of the encoding matrix in parallel. GPU implements the algorithm 83.33% faster than CPU.

[0048] FIGS. 5A, 5B, and 5C show the reconstructed images using CPU and GPU of simulated brain dataset using the proposed method. The time taken by CPU (Core 2 Duo 3.00 GHz) for simulated brain dataset is 191.3 ms while GPU (Nvidia GTX 560 Ti) for brain data set takes 19 ms to reconstruct the same image. GPU implementation of the proposed method is 10.06 times faster than CPU.

[0049] FIGS. 6A, 6B and 6C show the phantom dataset reconstructed images using QR based SENSE algorithm. FIG. 6A shows the original image, FIG. 6B shows the reconstructed image using CPU and FIG. 6C shows the reconstructed image using GPU. The time taken by CPU (Core 2 Duo 3.00 GHz) for phantom dataset is 814 ms while GPU (Nvidia GTX 560 Ti) for brain data set takes 192 ms to reconstruct the same image. GPU implements the algorithm 4.23 times faster than CPU.

[0050] FIGS. 7A, 7B, and 7C show the reconstructed images of cardiac dataset acquired using 30 channel receiver coils and reconstructed using CPU and GPU. FIG. 7A shows the original image, FIG. 7B shows the reconstructed image using CPU and FIG. 7C shows the reconstructed image using GPU. The time taken by CPU (Core 2 Duo 3.00 GHz) for 30 channel cardiac dataset is 3395 ms while GPU (Nvidia GTX 560 Ti) takes 217.87 ms to reconstruct the same image. GPU implements the algorithm 15.5 times faster than CPU.

[0051] Table 1 shows the comparison of time taken by CPU and GPU for the reconstruction of simulated brain, phantom and cardiac datasets. A comparison of the artifact power (AP) is also given in table 1 which shows that the AP is same for both CPU and GPU implementations.

TABLE-US-00001 TABLE 1 Comparison of the time taken by CPU and GPU for Brain, Phantom and Cardiac datasets. Reconstruction Dataset No. of Time (ms) Artifact Power used Coils AF CPU GPU Speedup (AP) Simulated 8 2 191.3 19 10.06x 0.0041 Brain Phantom 8 2 814 192 4.23x 0.00091 Cardiac 30 2 3395 217.87 15.5x 0.0057

[0052] FIG. 8 shows the comparison of time taken by CPU and GPU for the 8 receiver coil simulated human head and phantom. The results show that GPU takes less time to compute inversion of the encoding matrix as compared to CPU without any degradation in the quality of the image.

[0053] FIG. 9 gives a comparison of the time taken by CPU and GPU for 30 channel cardiac dataset. As shown by the bar graph, GPU takes less time to compute inversion of the encoding matrix as compared to CPU without any degradation in the quality of the image.

[0054] 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.