DCE-MRI has become the reference technique to detect and characterize breast lesions. A standard protocol, as described in clinical guidelines¹, consists in six 3D T1-eighted Fast Spoiled Gradient Recalled Echo (FSPGR) acquisitions covering the whole chest during a contrast agent (CA) injection, with high spatial resolution (below 1 mm) and a temporal resolution around 90 sec. A fat suppression is also desired to enhance contrast between lesions and fatty tissue, which can be implemented efficiently using a spectral inversion preparation (SPECIAL/SPIR). Despite the use of conventional acceleration techniques (SENSE, Partial Fourier), the duration of one acquisition remains quite long. To obtain more information about lesion vascularization, new acceleration techniques have emerged including TRICKS/TWIST² and Compressed Sensing (CS)³.

In this study, we propose a k-space trajectory allowing, from a single acquisition dataset (i.e. from six fully sampled k-spaces with 3D FSPGR), a flexible reconstruction that can provide both: i) the standard images that fulfill all requirements of the clinical guidelines; ii) high-temporal resolution data based on CS acquisition and reconstruction technique.The design is based on a smart reordering of the k-space lines so that each full k-space can be retrospectively split into a chosen number of frames with random but complementary sampling. Importantly the smart reordering needs to take account of the fat suppression scheme.

MRI protocol: MRI experiments were performed on a 3T General Electric scanner in one volunteer using an 8-channel breast coil. In this study we used the same sequence as in the clinical protocol for breast DCE-MRI in our center: 3D-FSPGR, 224x244x128 matrix, “SPECIAL” fat saturation. Only the k-space sampling was changed. Efficient fat suppression is achieved by using a spectrally selective chemical saturation technique. An inversion pulse centered on the fat frequency is played out every N recorded data. To guarantee a good fat suppression, the most central lines of the Fourier domain need to be acquired when the fat signal is minimal between two saturation pulses.

K-space trajectory design: To achieve a random sampling of the Fourier domain in both phase and slice directions, a random permutation of (ky,kz) pairs (ky from 1 to 244 and kz from 1 to 128) is first generated. The fully acquired Fourier domain may now be divided into as many frames as desired (Figure 1: example with 4 frames – acceleration rate of the protocol = 4). Then, within each frame, the (ky,kz) pairs are sorted according to their distance to the central kz plane (as it is done in the conventional sequence) and separated into N groups (Figure 2 : example with N=5 ). After a given fat inversion pulse is applied, (ky,kz) pairs are picked up in groups 4-2-1-3-5 (with 1 being the group of most central k-space lines) as shown in Figure 3.

The proposed sampling scheme for random Fourier domain subsampling allowed a good fat saturation. Figure 4 shows reconstructed breast images acquired in a volunteer. Fat aliasing artifacts spreading into phase and slice directions can be seen when a naïve random sampling scheme is applied. Conversely, no fat artifacts can be seen when the proposed reordering is applied, as shown in Figure 5.

The smart k-space trajectory design makes random k-space acquisition techniques compatible with spectral fat saturation. The proposed protocol allows a flexible use of the acquired data at the reconstruction stage. On one hand, we can choose to reconstruct the images from the fully sampled k-space data, as described in clinical guidelines for DCE-MRI of breast. One the other hand, reconstructed high-temporal resolution dataset can be reconstructed using CS after a simple splitting of the k-space into a chosen number of frames.

In future work, CS reconstructions will be tested in non-injected volunteers, injection phantom specially designed for the study and in injected patients as part of a clinical protocol.