3D mobile regression vision transformer for collateral imaging in acute ischemic stroke

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate collateral perfusion assessment is vital for acute ischemic stroke diagnosis and treatment.
• Current imaging methods (CT angiography, MR perfusion/angiography) are time-consuming or suboptimal.
• Deep learning shows promise but faces computational challenges.

Purpose of the Study:

• To develop a mobile, lightweight deep regression neural network for collateral imaging in acute ischemic stroke.
• To balance model complexity and performance for efficient collateral assessment.

Main Methods:

• Utilized dynamic susceptibility contrast MR perfusion (DSC-MRP) data from 952 patients.
• Developed a novel neural network integrating lightweight convolution and Transformer architectures.
• Generated five-phase collateral maps (arterial, capillary, early venous, late venous, delayed).

Main Results:

• The proposed lightweight model outperformed comparable deep learning models.
• Achieved performance comparable to more complex deep learning approaches.
• Demonstrated effectiveness in generating multi-phase collateral maps.

Conclusions:

• The developed model enables rapid and precise collateral status assessment in acute ischemic stroke.
• Facilitates improved patient care and outcomes through efficient imaging analysis.
• Represents a significant advancement in AI-driven neurological diagnostics.