Deep Learning-Based Image Enhancement of Low-Field Brain MRI: A Multicenter Validation of Diagnostic Image Quality and Lesion Detection

Abstract

Background: Low-field MRI (LF-MRI) widens neuroimaging access in resource-limited settings but suffers low signal-to-noise ratio (SNR), reduced resolution and artefacts. We developed and validated a deep-learning framework for image normalisation and noise reduction to elevate 0.35T brain MRI toward high-field quality, and tested whether it improves clinically significant lesion detection. Methods: In a multicentre retrospective diagnostic-accuracy study (STARD 2015), 450 adults underwent non-contrast 0.35T brain MRI (T1W, T2W, FLAIR) across three private tertiary centres in Palembang, Indonesia. Images were enhanced with a CycleGAN incorporating Vision-Transformer blocks. Three blinded neuroradiologists scored a 5-point Likert scale and recorded lesion presence; paired 1.5T MRI was the reference standard. Sensitivity, specificity, AUC and likelihood ratios were computed with 95% CIs; tests compared by McNemar and DeLong; agreement by Fleiss kappa. Results: AI enhancement improved all quality metrics (e.g., T1W PSNR 22.15 to 28.45 dB; SSIM 0.71 to 0.89; all p<0.001). For lesion detection, AI-enhanced LF-MRI achieved sensitivity 93.9% (95% CI 89.4-96.6), specificity 91.1% (87.1-94.0) and AUC 0.94 (0.91-0.97) versus 78.3%, 81.1% and 0.81 for original images (DeLong p<0.001; McNemar p<0.001). LR+ rose to 10.56 and LR- fell to 0.067. Inter-reader agreement was almost perfect (Fleiss kappa 0.78-0.85). Conclusions: A CycleGAN-with-transformer framework substantially improved objective quality and diagnostic performance of 0.35T brain MRI toward high-field standards with almost-perfect reader agreement. Pending prospective and external validation, AI enhancement is a low-cost route to more equitable neuroimaging.