Ronnie Wirestam

PURPOSE: Four-pool Voigt (FPV) machine learning (ML)-based fitting for Z-spectra was developed to reduce fitting times for clinical feasibility in terms of on-scanner analysis and to promote larger cohort studies. The approach was compared to four-pool Lorentzian (FPL)-ML-based modeling to empirically verify the advantage of Voigt models for Z-spectra.

METHODS: Voigt and Lorentzian models were fitted to human 3 T Z-spectral data using least squares (LS) to generate training data for the corresponding ML versions. Gradient boosting decision trees were trained, resulting in one Voigt and one Lorentzian ML model. Modeling accuracy was tested, and the fitting times of the ML models and LS versions were evaluated. The goodness of fits of Voigt and Lorentzian ML models were compared.

RESULTS: The training time for each ML model (Voigt and Lorentzian) was less than 1 min, and the modeling accuracy compared to the corresponding LS versions was excellent, as indicated by a nonsignificant difference between the parameters obtained by LS and corresponding ML versions. The average fitting time was 20 μs/spectrum for both ML models compared to 0.27 and 0.82 s/spectrum for LS with FPL and FPV, respectively. The goodness of fits of FPV-ML and FPL-ML differed significantly (p < 0.005), with FPV-ML showing an improvement for all tested data.

CONCLUSION: Gradient boosting decision trees fitting of multi-pool Z-spectra significantly reduces fitting times compared to traditional LS approaches, allowing fast data processing while upholding fitting quality. Along with the short training times, this makes the method suitable for clinical settings and for large cohort research applications. The FPV-ML approach provides a significant improvement of goodness of fit compared to FPL-ML.