AbstractPurposeProton (1H)‐MRSI via spatial‐spectral encoding poses high demands on gradient hardware at ultra‐high fields and high‐resolutions. Rosette trajectories help alleviate these problems, but at reduced SNR‐efficiency because of their k‐space densities not matching any desired k‐space filter. We propose modified rosette trajectories, which more closely match a Hamming filter, and thereby improve SNR performance while still staying within gradient hardware limitations and without prolonging scan time.MethodsAnalytical and synthetic simulations were validated with phantom and in vivo measurements at 7 T. The rosette and modified rosette trajectories were measured in five healthy volunteers in 6 min in a 2D slice in the brain. An elliptical phase‐encoding sequence was measured in one volunteer in 22 min, and a 3D sequence was measured in one volunteer within 19 min. The SNR per‐unit‐time, linewidth, Cramer‐Rao lower bounds (CRLBs), lipid contamination, and data quality of the proposed modified rosette trajectory were compared to the rosette trajectory.ResultsUsing the modified rosette trajectories, an improved k‐space weighting function was achieved resulting in an SNR per‐unit‐time increase of up to 12% compared to rosette's and 23% compared to elliptical phase‐encoding, dependent on the two additional trajectory parameters. Similar results were achieved for the theoretical SNR calculation based on k‐space densities, as well as when using the pseudo‐replica method for simulated, in vivo, and phantom data. The CRLBs of γ‐aminobutyric acid and N‐acetylaspartylglutamate improved non‐significantly for the modified rosette trajectory, whereas the linewidths and lipid contamination remained similar.ConclusionBy optimizing the shape of the rosette trajectory, the modified rosette trajectories achieved higher SNR per‐unit‐time and enhanced data quality at the same scan time.