Improved myelin water quantification using spatially regularized non‐negative least squares algorithm

Abstract

Purpose: To improve the myelin water quantification in the brain in the presence of measurement noise and to increase the visibility of small focal lesions in myelin-waterfraction (MWF) maps. Materials and Methods:A spatially regularized non-negative least squares (srNNLS) algorithm was developed for robust myelin water quantification in the brain. The regularization for the conventional NNLS algorithm was expanded into the spatial domain in addition to the spectral domain. Synthetic data simulations were performed to study the effectiveness of this new algorithm. Experimental free-induction-decay measurements were obtained using a multi-gradient-echo pulse sequence and MWF maps were estimated using the srNNLS algorithm. The results were compared with other conventional methods.Results: A substantial decrease in MWF variability was observed in both simulations and experimental data when the srNNLS algorithm was applied. As a result, false lesions in the MWF maps disappeared and the visibility of small focal lesions improved greatly. On average, the contrast-tonoise ratio for focal lesions was improved by a factor of 2. Conclusion:The MWF variability due to the measurement noise can be substantially reduced and the detection of small focal lesions can be improved by using the srNNLS algorithm.Key Words: myelin content; myelin water fraction; multiple sclerosis; non-negative least squares; regularization; multiexponential analysis. QUANTITATIVE MEASUREMENTS of myelin content can substantially improve our understanding of the pathological progress of several white matter (WM) diseases, such as multiple sclerosis (MS) (1-4). A technique that provides specific and sensitive information about the myelin content was developed based on an analysis of T 2 relaxation times (5,6). It has been reported that the T 2 spectrum of WM and several myelinated tissue samples consists of multiple components (7-13) and that the short component with T 2 between 10 ms and 50 ms corresponds to the water pool within the myelin sheath (2-6). Myelin content can therefore be quantitated using the fraction of this short T 2 component. To implement this method, T 2 decay signals are acquired using a 32-echo single-slice Carr-Purcell-Meiboom-Gill (CPMG) sequence with composite 90x-180y-90x refocusing pulses and big crusher gradients around the refocusing pulses (5,6,14,15). A non-negative least squares (NNLS) algorithm can then be used to estimate the T 2 spectrum from the acquired decay signal, and the myelin water fraction (MWF) can be calculated from the ratio of the short T 2 component (10 ms Ͻ T 2 Ͻ 50 ms) to the total (16). A strong correlation was found between the MWF measured using this technique and the myelin distribution using histopathology in fixed brains (4,17). This technique has been used to quantitatively measure the MWF in brains of subjects with MS (3-6,17).T 2 spectrum regularization has been applied to the NNLS algorithm to improve the reliability of fitting in the presence of noise. The regularized non-negative least squ...

In vivo multi-slice mapping of myelin water content using T2* decay

“…An adequate SNR was achieved with the use of an 8-channel phased-array coil and an anisotropic diffusion filter in a single acquisition at 3 T. High-resolution and multi-slice MWF maps were obtained in a clinically acceptable scan time. Du, 2009b). Since our three-pool model consists of discrete components, the use of a smoothing constraint is not suitable.…”

Multi-Compartment T2 Relaxometry Using a Spatially Constrained Multi-Gaussian Model

“…Spatially smoothed versions of conventionally obtained T2/T2* distributions were proposed as “reference distributions” and deviations from this was penalized [41] [42]. These approaches do not amount to a truly spatially constrained approach, because the inverse problem is still solved voxel-wise.…”

High-quality multiple T2(*) contrast MR images from low-quality multi-echo images using temporal-domain denoising methods

“…Since the decay rate of the MR temporal signal at each location is governed by MR physics known as T 2 (*) relaxation, the decay signal can be modeled with an exponential function. 12,13,16 If a single voxel contains different tissues, then the decay signal can be modeled with a multi-exponential function. 14,15 Therefore, the decay signal was fit into a specific multi-exponential function by a minimization process, such as a least squares algorithm.…”

“…Myelin was assumed to be damaged. 16,17 Gaussian random noise was added to decay signals. The SNR of the noisy simulation data was about 80 dB at the first echo time.…”