Perfusion parameter map generation from TOF-MRA in stroke using generative adversarial networks.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Cerebrovascular Diseases

Background:

• Traditional perfusion imaging techniques can be invasive or time-consuming.
• Time-of-flight magnetic resonance angiography (TOF-MRA) is a non-invasive imaging modality.
• Developing non-invasive methods to assess cerebral hemodynamics is crucial for cerebrovascular disease management.

Purpose of the Study:

• To develop an AI model capable of generating perfusion parameter maps from TOF-MRA images.
• To provide a non-invasive alternative to conventional perfusion imaging methods.
• To assess cerebral hemodynamics in patients with cerebrovascular diseases.

Main Methods:

• A retrospective study included 272 patients with cerebrovascular diseases (acute stroke and steno-occlusive disease).
• A 3D pix2pix generative adversarial network (GAN) was adapted to generate perfusion maps (CBF, CBV, MTT, TTP, Tmax) from TOF-MRA images.
• Model performance was evaluated using structural similarity index measure (SSIM) and Dice coefficient for lesion overlap.

Main Results:

• The GAN model demonstrated high visual overlap and performance across all generated perfusion maps.
• Quantitative metrics (SSIM, PSNR, MAE, NRMSE) indicated robust performance in both patient cohorts.
• Lesion overlap analysis showed a median Dice coefficient of 0.49 for hypoperfused lesions (Tmax > 6 s).

Conclusions:

• The AI model successfully generates perfusion maps from TOF-MRA, offering a non-invasive assessment of cerebral hemodynamics.
• This AI-driven approach has the potential to impact patient stratification in cerebrovascular diseases.
• Further refinement and validation are warranted for widespread clinical application.