Modeling Ischemic Stroke Pathological Dynamics via Continuous Fields and Vector Flow.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate localization of perfusion deficits in diffusion-weighted MRI (DWI) is crucial for managing acute ischemic stroke.
• Current deep learning models generate discrete binary masks, losing critical intra-lesion information and failing to represent the continuous nature of ischemic injury.

Purpose of the Study:

• To introduce StrokeFlow, a novel framework for representing ischemic regions as continuous fields.
• To develop a coordinate-based network generating a smooth ischemic density field for voxel-level infarction probability.
• To incorporate a vector flow head for modeling perfusion deficit directionality using the Apparent Diffusion Coefficient (ADC) map gradient.

Main Methods:

• Developed a novel coordinate-based deep learning network (StrokeFlow).
• Trained the network to output a continuous ischemic density field.
• Implemented a vector flow head supervised by the negative gradient of the ADC map.

Main Results:

• StrokeFlow achieved superior lesion boundary accuracy, significantly outperforming baselines in the 95% Hausdorff Distance on the ISLES 2022 dataset.
• The model demonstrated enhanced sensitivity for detecting small and multifocal ischemic lesions.
• Continuous field representation provided more biologically plausible and interpretable results compared to discrete segmentation.

Conclusions:

• StrokeFlow shifts the paradigm from discrete segmentation to continuous, functionally-aware fields for ischemic stroke assessment.
• The framework offers a more nuanced and clinically valuable tool for evaluating perfusion deficits in acute ischemic stroke.
• This approach enhances the precision and interpretability of MRI-based stroke analysis.