Deep Learning-Based Synthetic TOF-MRA Generation Using Time-Resolved MRA in Fast Stroke Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Time-resolved MRA is crucial for evaluating collateral circulation in acute ischemic stroke with large-vessel occlusion.
• Limitations in signal-to-noise ratio (SNR) and spatial resolution of time-resolved MRA hinder accurate diagnosis of vascular occlusion.
• Developing advanced imaging techniques is essential to overcome these diagnostic challenges.

Purpose of the Study:

• To develop and evaluate a CycleGAN-based deep learning model for generating high-resolution synthetic Time-of-Flight MRA (TOF-MRA) images from time-resolved MRA.
• To assess the image quality and clinical efficacy of these synthetic TOF-MRA images.

Main Methods:

• A retrospective study included 397 patients undergoing both TOF-MRA and time-resolved MRA.
• A CycleGAN deep learning model was trained on a subset of patients and validated on another.
• Image quality was assessed qualitatively and quantitatively, alongside a multireader diagnostic evaluation and clinical validation in acute ischemic stroke cases.

Main Results:

• Synthetic TOF-MRA showed improved image quality metrics (overall quality, sharpness, SNR) compared to time-resolved MRA for specific arterial segments.
• Radiologists could not distinguish synthetic TOF-MRA from real TOF-MRA in a blinded evaluation.
• Clinical validation demonstrated increased diagnostic confidence and reduced decision time for large-vessel occlusion detection.

Conclusions:

• A CycleGAN-based deep learning model effectively generates synthetic TOF-MRA from time-resolved MRA.
• Synthetic TOF-MRA holds potential to aid in the detection of large-vessel occlusion in stroke centers utilizing time-resolved MRA.