A Gradient Boosted Regression Tree Ensemble Model Using Wavelet Features for Post-acquisition Macromolecular Baseline Isolation from Brain MR Spectra

Abstract

Broad macromolecular baseline (MMBL) is present throughout the magnetic resonance spectroscopy (MRS) spectrum of brain at short echo-time (TE) acquisitions. These variations cause the metabolite peak quantification and processing difficult in diagnostics. MMBL can provide information for specific disease, as an important biomarker. This presents a requirement of an efficient MMBL isolation method in post-acquisition scenario. The volume of medical dataset available are mostly small- or medium-sized along with the constraint of ground truth. The estimation of MM baseline from a noisy spectrum was treated as an ill-conditioned inverse problem. To address both issues, a novel approach of gradient boosted wavelet-feature tree model in a multioutput-regression framework for MRS spectral fitting was adopted to isolate macromolecular baseline from noisy metabolite spectra, where the inverse problem was learned by training over wavelet coefficients of noisy spectral dataset simulated using basis-set of metabolites and macromolecules. The proposed method performed almost perfectly for the simulated dataset with smaller margins of error, compared to an equivalent CNN model. For the simulated test set, RMSE and SSIM of 0.1623 and 0.9571 respectively were obtained and RMSE of 0.2263 was obtained for in-vivo test set. The fitted peak amplitude for individual MM component within ± 4% of error range over the simulated dataset. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Austria, part of Springer Nature.