Evaluation of generative adversarial network-based postprocessing super-resolution for lumbar spine magnetic resonance imaging

This study evaluated the feasibility of generative adversarial network (GAN)-based postprocessing super-resolution for T2-weighted lumbar spine magnetic resonance imaging (MRI) using both objective resolution metrics and perceptual assessment. Sagittal lumbar spine MRI datasets from healthy volunteers were analyzed. An enhanced super-resolution GAN (ESRGAN) was trained on downsampled images, while zero-filling, bicubic, and bilinear interpolation, as well as enhanced deep residual networks for single image super-resolution (EDSR), were used as comparators. Image quality was assessed by comparing upscaled images from 256 × 256 acquisitions with corresponding 512 × 512 reference images, using full width at half maximum (FWHM) and normalized integrated power spectrum (NIPS), along with similarity metrics including structural similarity index, peak signal-to-noise ratio, and root mean square error. Subjective image quality was evaluated using a ranking method. An additional analysis using 320 × 320 and 640 × 640 image pairs was conducted to assess consistency across resolution settings. Statistical comparisons were performed using the Friedman test with Bonferroni correction. ESRGAN showed no significant differences from the original high-resolution images in FWHM and NIPS, whereas interpolation methods and EDSR demonstrated inferior performance (P < 0.05). Although interpolation methods achieved higher scores in pixel-wise metrics, ESRGAN obtained the highest subjective ratings and interrater agreement. These findings indicate that ESRGAN better restores high-frequency structural information not captured by conventional similarity metrics. GAN-based postprocessing super-resolution may improve image quality in lumbar spine MRI and warrants further investigation of its clinical impact.