Gray matter (GM) imaging has gained increasing attention in the studies of brain disease (e.g., multiple sclerosis¹) using double inversion recovery (DIR) pulse sequence². However, both long and short inversion times (TIs) are required in conventional DIR-based GM imaging, substantially prolonging imaging time while reducing the signal intensity of GM. Thus, it has not been feasible to achieve high resolution whole-brain GM imaging within a clinically feasible imaging time. In this work, we propose a novel, GM-selective single-slab 3D dual-echo TSE with relaxation modulation (no long IR), in which white matter (WM) signals remain nulled before signal encoding while CSF signals are modulated to be cancelled during residual reconstruction between echoes with sparsity prior. We demonstrate the effectiveness of the proposed method over conventional DIR in that the former can achieve 1mm-isotropic whole-brain GM-images in 5-6 min. without apparent artifacts.In conventional DIR¹ (Fig. 1a), both long and short IR preparations are applied to attenuate CSF and WM signals, respectively, while variable flip angles (VFA) are employed for a three-step (exponential-flat-exponential) signal evolution of GM signals along the echo train. In the proposed method (Figs. 1b,c), the first half of the echo train (ECHO₁) acquires short IR prepared, WM-suppressed signals with VFA for a two-step (flat-exponential) GM signal evolution while the second half of the echo train (ECHO₂) CSF-dominant signals with linearly decreasing refocusing flip angles for a smooth transition of pseudo-steady-state (PSS) along the echo train³.Incoherent sparse undersampling in an elliptical k_y-k_z-space was employed. To maximize GM signals while balancing CSF signals over the two ECHOes, signals along the echo train are filled in corresponding ky-kz-space with centric (center in->out) and reverse centric (center out->in) reordering fashions in ECHO₁ and ECHO₂, respectively. To be robust to noise amplification during image-based weighted averaging, in this work GM images are reconstructed directly from the signal difference between the two k-spaces for ECHO₁ (y₁) and ECHO₂ (y₂) (Δy=y₁-ω·y₂; ω:weighting parameter) by solving the following optimization problem: $$$\arg \underset{\Delta \mathbf{x}}{\operatorname{min}} \; \| \mathbf{F_uSx}-\Delta\mathbf{y}\|^2_2+\lambda\|\Psi\Delta\mathbf{x}\|_1$$$ where $$$\Delta\mathbf{x}$$$ is the GM image, $$$\mathbf{S}$$$ is the coil sensitivity matrix, $$$\mathbf{F_u}$$$ is the undersampled Fourier transform, $$$\lambda$$$ is the regularization parameter, and $$$\Psi$$$ is the sparsifying transform. The fully sampled, inner k-space data was employed to calculate ω as well as coil sensitivity map. To investigate refocusing flip angles and corresponding signal evolutions of brain tissues and CSF along the echo train in conventional DIR and the proposed method, numerical simulations of Bloch equation were performed using the parameters: ESP=3.3ms, Short TI=550ms, maximum FA=160˚, those specific to conventional DIR: TR=10s, Long TI=3100ms, and echo train length (ETL)=200; and those specific to the proposed method: TR=4s, ETL=240, and last FA=50˚.Experiments were performed in two healthy volunteers on a 3T (Magnetom Trio, Siemens Medical Solutions, Erlangen, Germany) using conventional DIR and the proposed method. A 32-channel head coil was used for signal reception. The imaging parameters common to both methods were: FOV= 250x204mm², matrix size=256x200, partitions=176, thickness=1mm, and bandwidth=781Hz/pix; the remaining parameters were identical to those above. The imaging time was 30min. in conventional DIR and 6min. in the proposed method. Additionally, three sets of GM image data were acquired with holding imaging time to 5.5min: 1) conventional DIR with a large slice thickness (5mm), 2) conventional DIR with high undersampling and GRAPPA reconstruction (acceleration factor: 2(k_y)X4(k_z)), and 3) the proposed method, and then qualitatively compared. Figure 2 shows the refocusing flip angles (Figs. 2a,c) and the corresponding signal evolutions of GM, WM, and CSF along the echo train (Figs. 2b,d) in conventional DIR (Figs. 2a,c) and the proposed method (Figs. 2b,d). The signal intensity of GM for the k-space center in the proposed method is much higher than that in conventional DIR. In the proposed method, CSF signals in the early and late portion of the echo train, which correspond to the k-space center in ECHO₁ and ECHO₂, respectively, are approximately equal, and thus leading to ω close to one. Figure 3 shows GM images in three orthogonal orientations in conventional DIR (Fig. 3a) and the proposed method (Fig. 3b). The proposed method yields a similar level of SNR despite much reduced imaging time. When the imaging time in conventional DIR and the proposed method is set identical to each other, the former suffers from severe blurring (Fig. 4a) or artifacts (Fig. 4b) whereas the latter produces 1mm-isotropic whole-brain GM images (Fig. 4c). In conclusion, the proposed method can be a promising alternative to conventional DIR, and is expected to widen its application to brain disease studies.