Speckle pattern analysis with deep learning for low-cost stroke detection: a phantom-based feasibility study.

Significance: Stroke is a leading cause of disability worldwide, necessitating rapid and accurate diagnosis to limit irreversible brain damage. However, many advanced imaging modalities (computerized tomography, magnetic resonance imaging) remain inaccessible in remote or resource-constrained settings due to high costs and logistical barriers. Aim: We aim to evaluate the feasibility of a laser speckle-based technique, coupled with deep learning, for detecting simulated stroke conditions in a tissue phantom. We investigate whether speckle patterns can be leveraged to differentiate healthy from restricted flow states in arteries of varying diameters and depths. Approach: Artificial arteries (3 to 6 mm diameters) were embedded at different depths (0 to 10 mm) within a skin-covered chicken tissue, to mimic blood-flow scenarios ranging from no flow (full occlusion) to high flow. A high-speed camera captured the secondary speckle patterns generated by laser illumination. These video sequences were fed into a three-dimensional convolutional neural network (X3D_M) to classify four distinct flow conditions. Results: The proposed method showed high classification accuracy, reaching 95% to 100% for larger vessels near the surface. Even for smaller or deeper arteries, detection remained robust ( Conclusions: Although tested on a tissue phantom, these findings highlight the potential of combining speckle imaging with deep learning for accessible, rapid stroke detection. Our next steps involve direct