Abstract
BACKGROUND: Spectroscopic MRI (sMRI) is a quantitative imaging technique that maps infiltrated tumors in the brain without contrast injections. In a previous study (NCT03137888), sMRI-guided radiation treatment extended patient survival, showing promise for clinical translation. The spectral fitting of individual voxels in an sMRI dataset generate metabolite concentration maps that guide treatment. The established spectral analysis methods use iterative least-squares fitting (FITT) that are computationally demanding. This study compares the performance of NNFit, a neural network-based, accelerated spectral fitting model, to the established FITT for metabolite quantification and radiation treatment planning. METHODS: NNFit is a self-supervised deep learning model trained on 50 ms echo-time (TE) sMRI data to estimate metabolite levels of choline (Cho), creatine (Cr), and NAA. We trained the model on 30 GBM patients (56 scans) and tested it on 17 GBM patients (29 scans). NNFit's performance was compared to the FITT using structural similarity indices (SSIM) and the Dice coefficient. RESULTS: NNFit significantly improved processing speed while maintaining strong agreement with FITT. The radiation target volumes defined by Cho/NAA ≥ 2x were visually comparable, with fewer artifacts in NNFit. Structural similarity indices (SSIM) indicated minimal bias and high consistency across methods. CONCLUSIONS: This study highlights NNFit's potential for rapid, accurate, and artifact-reduced metabolic imaging, enabling faster radiotherapy planning.