GABA+-Edited Magnetic Resonance Spectroscopy Deep Learning Quality Assessment Framework.
Authors
Affiliations (5)
Affiliations (5)
- Biomedical Engineering Department, University of Calgary, Calgary, Canada.
- Department of Radiology, University of Calgary, Calgary, Canada.
- Hotchkiss Brain Institute, University of Calgary, Calgary, Canada.
- Alberta Children's Hospital Research Institute, University of Calgary, Calgary, Canada.
- Electrical and Software Engineering Department, University of Calgary, Calgary, Canada.
Abstract
Motivated by the need to improve GABA+-edited magnetic resonance spectroscopy (MRS) quality, we developed a three-module framework to improve transient averaging based on quality. We hypothesized that training a deep learning (DL) model to differentiate spectrum quality could improve transient averaging compared to traditional averaging (simple arithmetic known as Equal-weighting) and an existing software weighting algorithm (MSE-weighting). The transient averaging framework was approached through three modules: (1) a continuous-valued automated quality labeling algorithm using both traditional and recently developed MRS quality metrics, (2) a dual-domain (time and frequency) DL model that learns from these quality labels to assess quality scores for new data, and (3) a transient weighting algorithm informed by DL quality scores. The labeling algorithm was used to produce quality labels focused on retaining GABA+ peak shape in difference spectra (1) to train the DL model (2). The DL model quality scores were used to assign weights (3) for transient pairs within the final average difference spectrum. Results were compared to MSE-weighting and Equal-weighting. Retaining only GABA+-edited transient pairs with positive quality labels resulted in improved spectral quality as assessed using traditional metrics, recently developed metrics and visual assessment of GABA+ and Glx peaks. Applying the trained DL model to in vivo scans improved the fit quality (lower fit error) compared to Equal-weighting (4.759 ± 1.545 vs. 4.877 ± 1.762) and produced higher SNR compared to MSE-weighting (18.758 ± 2.392 vs. 18.004 ± 2.68). The framework as proposed can moderately improve data quality by optimizing transient averaging and provides opportunties for new work.