Background: Portable low-field MRI (pMRI) offers rapid, bedside neuroimaging in emergency stroke unit (ESU) but suffering from low signal-to-noise ratio and blurred anatomy. We developed a residual diffusion model conditioned on original low-field images to enhance DWI, T2-FLAIR and T1w for acute stroke evaluation. Methods: We retrospectively analyzed 1,325 consecutive patients with suspected acute ischemic stroke imaged within 24 hours of onset using both 0.23 T pMRI and 3.0 T reference MRI. The model employs a Swin-UNet backbone with conditional feature modulation from the corresponding low-field scans, progressively suppressing noise and restoring fine structures across diffusion timesteps. Image quality was quantified by peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) against 3.0 T references. Segmentation-based clinical relevance was assessed using modality-specific tools: pre-trained nnUNet for lesion segmentation on DWI, WMH-SynthSeg for white-matter hyperintensity (WMH) evaluation on T2-FLAIR, and FreeSurfer for tissue-specific volumetry (white matter, gray matter, cerebrospinal fluid) on T1w. Inference was performed on an NVIDIA RTX 3090 GPU. Results: The model consistently improved fidelity across modalities. For DWI, PSNR increased from 19.1 to 26.1 dB and SSIM from 0.69 to 0.83, with lesion-volume correlation rising from r = 0.72 to r = 0.80. For T2-FLAIR, PSNR improved from 20.8 to 26.4 dB and SSIM from 0.66 to 0.85, with WMH-volume correlation from r = 0.81 to r = 0.88. For T1w, PSNR rose from 21.9 to 28.7 dB and SSIM from 0.74 to 0.90; correlations with 3.0 T volumes strengthened for WM (r = 0.479→0.936), GM (r = 0.390→0.945), and CSF (r = 0.726→0.767). Mean processing time was 3.7 s per scan. Conclusions: Conditioning a residual diffusion model on the native low-field image substantially improves pMRI quality and volumetric agreement with 3.0 T references across DWI, T2-FLAIR, and T1w, supporting accurate and rapid bedside assessment in ESU and underscoring the clinical potential of portable low-field MRI in acute stroke workflows.
Zhou et al. (Thu,) studied this question.