A deep autoencoder for fast spectral-temporal fitting of dynamic deuterium metabolic imaging data at 7T.
Authors
Affiliations (7)
Affiliations (7)
- High Field MR Center, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria; Comprehensive Center for AI in Medicine (CAIM), Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria.
- Department of Electrical and Software Engineering, University of Calgary, 622 Collegiate Place, Calgary, Alberta, Canada.
- High Field MR Center, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria.
- Centre for Integrative Neuroimaging, FMRIB, Nuffield Department of Clinical Neurosciences, University of Oxford, Headley Way, Oxford, United Kingdom.
- Comprehensive Center for AI in Medicine (CAIM), Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria; Computational Imaging Research Lab, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria.
- High Field MR Center, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria; Comprehensive Center for AI in Medicine (CAIM), Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria; Christian Doppler Laboratory for MR Imaging Biomarkers (BIOMAK), Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria. Electronic address: [email protected].
- High Field MR Center, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria; Comprehensive Center for AI in Medicine (CAIM), Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria; Christian Doppler Laboratory for MR Imaging Biomarkers (BIOMAK), Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Lazarettgasse 14, Vienna, 1090, Austria.
Abstract
Deuterium metabolic imaging (DMI) is a non-invasive magnetic resonance spectroscopic imaging technique enabling in vivo mapping of glucose metabolism. Dynamic DMI provides time-resolved metabolite maps and allows spatially resolved fitting of metabolic models to capture metabolite concentration dynamics. However, conventional fitting tools often require long processing times for high-resolution datasets, limiting their practical utility. To address this bottleneck, we propose a deep autoencoder (DAE) for rapid spectral-temporal fitting of dynamic DMI data, supporting arbitrary parametric model constraints to describe metabolite concentration dynamics. The DAE was benchmarked against spectral-temporal fitting using FSL-MRS and LCModel. Fitting accuracy was evaluated on in vivo and synthetic whole-brain dynamic DMI data acquired at 7T using Bland-Altman metrics, Pearson correlation coefficients, structural similarity index measures, and root mean squared errors for both metabolite concentrations and model constraint parameters. The DAE achieved processing times of 0.29 ms per voxel, corresponding to an acceleration of more than three orders of magnitude compared to LCModel/FSL-MRS (0.55/0.65 s per voxel). On in vivo data, it showed excellent agreement with LCModel, and on synthetic data, it consistently outperformed both reference methods across all evaluation metrics. The proposed DAE enables accurate spectral-temporal fitting for whole-brain dynamic DMI scans within less than a second, matching or exceeding the performance of conventional state-of-the-art fitting methods. This makes it a promising tool for integration into efficient post-processing pipelines for research and clinical DMI workflows.