Modeling Ischemic Stroke Pathological Dynamics via Continuous Fields and Vector Flow

Abstract

Precise localization of perfusion deficits in diffusion-weighted MRI (DWI) is critical for acute ischemic stroke management. However, existing deep learning methods typically produce discrete binary masks, failing to capture the continuous nature of ischemic injury and discarding valuable intra-lesion information. We propose StrokeFlow, a novel framework that represents the ischemic region as a continuous field. Our coordinate-based network is trained to output a smooth ischemic density field, representing voxel-level infarction probability. Furthermore, we introduce a vector flow head, explicitly supervised to learn a vector field that aligns with the negative gradient of the Apparent Diffusion Coefficient (ADC) map, thereby modeling the directionality of the perfusion deficit. Evaluated on the public ISLES 2022 dataset, StrokeFlow demonstrated superior lesion boundary accuracy, significantly outperforming strong baselines in the 95% Hausdorff Distance metric. The model also showed enhanced sensitivity in detecting small and multifocal lesions. By shifting the paradigm from discrete segmentation to continuous, functionally-aware fields, StrokeFlow offers a more biologically plausible and interpretable tool for a nuanced clinical assessment of ischemic stroke.