The novel technique called as Time-of-flight with sparse undersampling and iterative reconstruction (Sparse TOF) has demonstrated the potential to accelerate TOF MRA ¹. Most of previous studies targeted healthy volunteers to develop algorithms or to optimize the parameters ², however, few clinical researches targeting patients have not been published. In this work, we conducted comparison study targeting patients with cerebral aneurysms to check the clinical relevance of evaluation of aneurysms on Sparse TOF and TOF with parallel imaging (PI TOF).

IRB approved this prospective study.

The 18 patients underwent non-contrast Sparse-TOF 3X, 5X, and these data were compared with non-contrast TOF with PI TOF (GRAPPA 3X. All images were obtained at 3T MR (MAGNETOM Skyra, Siemens, Erlangen, Germany) by using 32-channel head coil.

PI TOF

The imaging prameters of PI TOF are as follows: TR/TE, 20/3.69 msec, Flip angle, 18 degree, FOV, 202.5 × 240 mm, Matrix, 270 × 320, Slice thickness, 0.38 mm, GRAPPA 3. The images were automatically interpolated to 540 × 640 matrix, slice resolution, 0.38 × 0.38 mm.

Sparse TOF

The imaging parameters of Sparse TOF are as follows: TR/TE, 20/3.69 msec, Flip angle, 18 degree, FOV 202 × 220, Matrix 288 ×264, slice resolution, the images were automatically interpolated to 576 × 528 matrix, slice thicknes, 0.38 mm, slice resolution 0. 38 × 0.38 mm. Sparse TOF data was reconstructed using a non-linear iterative SENSE-based algorithm with a constraint to enforce sparsity. Specifically, images were reconstructed by solving the following minimization problem with a Modified Fast Iterative Shrinkage-Thresholding Algorithm (mFISTA).

(If you cannot see the equation, please refer to Figure 5)

where x is the image to reconstruct, yj and Sj are the k-space data and coil sensitivity for j-th coil element, Fu is the Fourier undersample operator, W is the redundant Haar wavelet transform, and lambda is the normalized regularixation weighting factor. Sampling pattern in k-space was designed based on the vairable density Poisson disc pattern. Undersampled data were seamlessly reconstructed on a standard reconstruction system, using 10 iterations with a regularization factor of 0.008. Total reconstruction was finished within 6 minutes.

Evaluation

One neuroradiologist blindly evaluated maximum intensity projection (MIP) images of without any prior information of the location of aneurysms. Aneurysms were graded with 3-point-scale: grade 3=excellently visible, grade 2=visible, grade 3=scarcely visible.

Patients

Twenty-three aneurysms were finally diagnosed by conventional angiography. Two aneurysms were found in 5 patients.

Examples of Image Reconstruction on Sparse TOF

Representative images of Sparse TOF 3X and 5X are shown in Figure 1. There are some aliasing artifacts and blurring artifacts in the source image of Sparse TOF with 1 iteration. With 10 iterataion, artifacts are reduced and more smoothed images are created. MIP images of representative cases are shown in Figure 2 and 3. There exist faint differences among PI TOF and Sparse TOF, however, the appearance is almost same and the visualization of aneurysms are sufficient for clinical diagnosis.

Evaluation

The sum of Grades evaluated by the neuroradiologist is shown in Figure 4. The scores are almost similar, however, the grade of Sparse TOF 3X was slightly better than that of PI-TOF and Sparse TOF 5X.

Sparse modeling is a very promising algorithm which is expected for effective undersampling in MR imaging. G-factor penalty is the unavoidable problem in parallel imaging, however, Sparse TOF provided us the clinically acceptable quality of MIP images even with 5X acceleration. Denoising effect may also benefit Sparse TOF 3X, because the sum of grade were slightly better than PI TOF 3X.Sparse TOF 3X and 5X were reconstructed with clinically acceptable time, and cerebral aneurysms were visible in both Sparse TOF 3X and 5X with equivalent quality as PI TOF.