Portable RF brain scanner enables stroke type detection using deep learning

Stroke remains a leading cause of mortality worldwide, demanding rapid, accurate diagnosis; distinguishing stroke types matters because treatments differ significantly. Using a recently developed, non-invasive brain scanner integrating a 16-antenna radio-frequency array, we present a deep-learning model that distinguishes the two main stroke types. We employ masked autoencoder-based self-supervised learning and supervised contrastive strategies to improve data efficiency and robustness with limited labeled clinical data. In the current test cohort, the system achieved 92% sensitivity and 85% specificity for hemorrhagic versus non-hemorrhagic detection, and 95% sensitivity and 80% specificity for ischemic versus non-ischemic cases. Beyond classification, the model showed patterns consistent with an ordering of relative dielectric permittivity across conditions (hemorrhagic, ischemic, mimic, and healthy). While not representing direct measurement of tissue dielectric properties, these observations provide insight into relative dielectric differences captured by the RF measurements. These findings support the potential of RF-based measurement-domain analysis to more reliable differentiation between stroke subtypes.