GPU-Accelerated Reconstruction of T2 Maps in Magnetic Resonance Imaging

Abstract

The main tissue parameters targeted by MR tomography include, among others, relaxation times T1 and T2. This paper focuses on the computation of the relaxation time T2 measured with the Spin Echo method, where the sensing coil of the tomograph provides a multi-echo signal. The maxima of these echoes must be interleaved with an exponential function, and the T2 relaxation can be determined directly from the exponential waveform. As this procedure needs to be repeated for each pixel of the scanned tissue, the processing of large images then becomes very intensive. For example, given the common resolution of 256x256 with 20 slices and five echoes at different times TE, it is necessary to reconstruct 1.3106 exponential functions. At present, such computation performed on a regular PC may last even several minutes. This paper introduces the results provided by accelerated computation based on parallelization and carried out with a graphics card. By using the simple method of linear regression, we obtain a processing time of less than 36 ms. Another effective option consists in the Levenberg-Marquardt algorithm, which enables us to reconstruct the same image in 96 ms. This period is at least 900 times shorter than that achievable with professional software. In this context, the paper also comprises an analysis of the results provided by the above-discussed techniques.The main tissue parameters targeted by MR tomography include, among others, relaxation times T1 and T2. This paper focuses on the computation of the relaxation time T2 measured with the Spin Echo method, where the sensing coil of the tomograph provides a multi-echo signal. The maxima of these echoes must be interleaved with an exponential function, and the T2 relaxation can be determined directly from the exponential waveform. As this procedure needs to be repeated for each pixel of the scanned tissue, the processing of large images then becomes very intensive. For example, given the common resolution of 256x256 with 20 slices and five echoes at different times TE, it is necessary to reconstruct 1.3106 exponential functions. At present, such computation performed on a regular PC may last even several minutes. This paper introduces the results provided by accelerated computation based on parallelization and carried out with a graphics card. By using the simple method of linear regression, we obtain a processing time of less than 36 ms. Another effective option consists in the Levenberg-Marquardt algorithm, which enables us to reconstruct the same image in 96 ms. This period is at least 900 times shorter than that achievable with professional software. In this context, the paper also comprises an analysis of the results provided by the above-discussed techniques.