Abstract WP272: AI-Enhanced Portable Low-Field MRI for Detecting Chronic Lacunes

Introduction: Cerebral small vessel disease (CSVD) is a vascular disorder associated with increased stroke risk and dementia. Neuroimaging markers of CSVD include small subcortical infarcts, white matter hyperintensities (WMH), microbleeds, enlarged perivascular spaces, atrophy, and lacunes. Portable, low-field (LF) MRI offers an accessible and affordable approach to CSVD assessment and longitudinal monitoring. While LF-MRI has been shown to detect WMH and regional atrophy, its sensitivity to detecting lacunes remains unknown. Advances in LF-MRI software integrate artificial intelligence (AI) to enhance resolution and improve image uniformity. We aimed to investigate the utility of LF-MRI for lacune evaluation and determine if optimized-AI software improves the accuracy of detection. Methods: Patients with suspected or a history of ischemic stroke were enrolled at Massachusetts General Hospital between August 2022 and July 2025. Participants underwent LF-MRI acquisition (0.064 T, Hyperfine Inc.) including T1, T2, and FLAIR sequences. Scans obtained after October 2024 used optimized-AI reconstruction software compared to prior software versions. Two assessors reviewed scans to identify the presence or absence of lacunes. The agreement, positive predictive value (PPV), negative predictive value (NPV), sensitivity, and specificity were calculated relative to ground-truth clinical MRI, reported as median and 95% CI. Results: A total of N =70 patients were enrolled (65±13 years, 58% female), of whom N =37 had confirmed lacunes on clinical MRI (7.0±2.4 mm). Of these patients, n =44 underwent LF-MRI on prior software and N =26 on optimized-AI software. Agreement between assessors was κ=0.53 (0.32-0.75) on prior software which increased to 0.63 (0.43-0.83) on optimized-AI. On prior software, the ability to detect lacunes was associated with a sensitivity of 44.4% (21.5-69.2), specificity of 100% (86.8-100), PPV of 100% (63.1-100), and NPV of 72.2% (54.8-85.8). On optimized-AI software, sensitivity increased to 83.3% (58.6-94.6; p<0.05), NPV to 72.7% (39-94, p>0.05), while specificity and PPV remained at 100% (63.1-100 and 78.2-100, respectively; p>0.05). Conclusions: We demonstrate that portable, LF-MRI can detect lacunar stroke, and when enhanced with AI-based reconstruction methods, improves the sensitivity of detection while maintaining specificity. These findings support the potential utility of LF-MRI for detecting chronic lacunes as part of CSVD assessment.