Super-resolution deep learning reconstruction for brain fluid-attenuated inversion recovery: image quality and white...

To evaluate whether super-resolution deep learning reconstruction (SR-DLR) improves two-dimensional (2D) brain fluid-attenuated inversion recovery (FLAIR) image quality while preserving automated white matter hyperintensity (WMH) volumetry, compared with Gaussian-filtered reconstruction (GA) and denoising DLR (dDLR).

Thirty-six healthy volunteers underwent 3T axial 2D FLAIR. Images were reconstructed using GA, dDLR, and SR-DLR (twofold in-plane upscaling). Quantitative metrics (noise, SNR, CNR, sharpness) were measured in standardized regions of interest. Two radiologists independently scored qualitative image quality. In a WMH-positive subgroup, WMH volumes were obtained using a transformer-based U-Net segmentation model on GA, dDLR, SR-DLR, and a high-resolution acquisition (HR-dDLR). Quantitative image-quality metrics were compared using one-way ANOVA, WMH volumes using paired t-tests, and interobserver agreement using weighted kappa.

SR-DLR demonstrated the lowest image noise and the highest SNR, CNR, and sharpness compared with GA and dDLR (all p < 0.001). Qualitative scores for noise, sharpness, and overall image quality were significantly higher for SR-DLR (p < 0.001), with perfect inter-observer agreement for sharpness and overall quality. Mean WMH volumes did not differ significantly across reconstructions, and SR-DLR volumes closely matched those of HR-dDLR.

SR-DLR substantially improves 2D brain FLAIR image quality at 3T by reducing noise and increasing SNR, CNR, and sharpness while preserving WMH volumetry consistent with high-resolution reference standards.