Spectral Subtraction Denoising (SSD) Improves Factor-3 GRAPPA 3T Lumbar MRI for Disc Herniation

Abstract

Parallel imaging shortens lumbar MRI examinations but factor-3 GRAPPA produces spatially varying noise that can obscure clinically important structures. Evidence for a simple, reproducible spectral subtraction denoising approach in clinically acquired 3T lumbar examinations remains limited. This paired pretest-posttest study evaluated MATLAB-based spectral subtraction denoising (SSD) in 16 patients with clinically diagnosed lumbar disc herniation. Sagittal T2-weighted TSE-GRAPPA images acquired at acceleration factor 3 were processed by estimating the background-noise spectrum, directly subtracting it from the image spectrum, retaining the original phase, and applying an inverse Fourier transform. Signal-to-noise ratio (SNR) was measured in five anatomical regions, and two radiologists with more than five years of experience independently performed blinded five-point visual grading. Paired differences were assessed with the Wilcoxon signed-rank test; acceleration-factor comparisons used the Friedman test, and interobserver agreement used Cohen’s kappa. SSD increased SNR in every evaluated structure, with the largest absolute gains in the vertebral body (+1.81) and degenerated nucleus pulposus (+0.41). Overall visual information improved (p < 0.001); four of five structures improved significantly, whereas the vertebral body showed a ceiling effect (p = 0.236). Interobserver agreement increased from κ = 0.60 to κ = 0.72. The findings support SSD as a low-complexity post-processing option for preserving diagnostic visibility in accelerated lumbar MRI. Larger multicenter studies should compare SSD with modern learning-based denoisers and evaluate diagnostic accuracy.