MRS data quality often depends on the skill level of the operator: whereas for expert users with compliant subjects it is possible to get excellent data even in challenging brain areas, studies acquired by real-world technologists without specific MRS training yield variable data, ultimately diminishing statistical power in research studies and diagnostic capability in the clinic. Current acquisitions proceed ‘blind’, with success judged by an expert during data analysis. The remarkable recent growth of artificial intelligence makes this approach archaic - we will therefore develop “intelligent” MRS sequences that self-monitor and adjust during acquisition to avoid data losses, such as from poor water suppression, field drift, pulsatile phase modulation, and instabilities. These adaptive acquisitions will be critical for edited MRS, which requires field stability, and MRS imaging (MRSI), where recognizing corrupted phase-encoding transients and using sparse reconstruction can prevent nuisance signals that propagate across the whole brain. Adaptive MRS requires rapid assessment of data quality – we will develop consensus quality control criteria, capturing expert knowledge and the relative weight given to each spectral feature (SNR, linewidth, artifacts, etc) in assessing quality, and validate that knowledge by modeling synthetic data of varying quality to ensure that sequences are adapting to spectral features that reduce modeling accuracy. 

Most of MRS and MRSI focuses on the metabolite proton signals a lower frequency than the water signal (the so-called upfield spectral region). At higher frequencies (downfield) relative to water, depending on the particular data acquisition method, one can assess aromatic and exchangeable proton signals.  We developed downfield MRSI (DF-MRSI) methodology both at 3T and 7T with whole brain coverage, evaluated its reproducibility and applied in brain cancers. We will now develop fast DF-MRSI acquisition and reconstruction methods using compressed sensing and deep learning.  The greatest biomarker potential of this technique is in amide proton mapping for assessing brain tumor malignancy, providing information on mobile cellular proteins, which can be detected with reproducibility at the clinical field strength of 3T.