Author: Anke Henning, Mahrshi Jani, Tianyu Wang, Andrew Wright, Xinyu Zhang 👨🔬
Affiliation: Advanced Imaging Research Center (AIRC), UT Southwestern Medical Center 🌍
Purpose: Proton MRSI offers critical metabolic insights into diseased brain processes but is prone to artifacts, and current post-processing methods are often insufficient, resulting in low-quality data. To address this gap, we present an unsupervised, dual-channel 1D-ResNet-based artifact removal method for clinical in-vivo 7T MRSI.
Methods: 1H-MRSI data from 40 subjects were categorized by artifact severity into clean (12), artifact-ridden (14), and severe-artifact groups (14) based on spectral pattern identifiability and artifact severity. Clean and artifact-ridden spectra, excluding skull and peripheral lipid signals, were randomly paired for training; severe-artifact group data was used for validation. Artifacts presented in spectra include gradient modulation sidebands, residual water (4.2–6 ppm), and lipid contamination (1–3 ppm). Spectra were zero-filled to 1024 points, normalized by NAA peak area (1.9, 2.1 ppm), and scaled to reduce inter-subject variability. A 1D-ResNet with 3 encoder-decoder level and Leaky ReLU activation was employed to reconstruct artifact-free spectra. Real and imaginary input components were concatenated to preserve relative relationships. To enhance metabolite peak shape and accuracy, a loss function with weighted L2 loss at metabolite range plus baseline-smoothness term was used. The model was trained with the Adam optimizer (batch size: 32) for 24 epochs.
Results: Reconstructed metabolite maps showed significant, consistent and robust artifact removal efficacy across various brain regions, enhanced peak shape for major metabolites (tCho, tCr, NAA) when compared to input; however, minor errors in peak height were observed. Quantitatively, average normalized RMSE for the residual water, lipid regions, and full spectrum decreased by 83.3%, 50%, and 70.6%, respectively. Sample metabolite maps revealed better concentration variation contrast.
Conclusion: We developed an effective artifact removal framework that substantially reduces artifacts in vivo 1H-MRSI. This method can salvage low-quality datasets and reduce repeat scans in clinical settings. Future studies will determine its impact on metabolite quantification accuracy.