Automated DWI-FLAIR mismatch assessment in stroke using DWI only.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Imaging

Background:

• Diffusion-Weighted Imaging (DWI) and Fluid-Attenuated Inversion-Recovery (FLAIR) mismatch is crucial for identifying Acute Ischemic Stroke (AIS) patients eligible for thrombolysis, especially with unknown onset times.
• Visual assessment of DWI-FLAIR mismatch suffers from observer variability, necessitating more objective methods.

Purpose of the Study:

• To develop and validate a Deep-Learning (DL) model for predicting DWI-FLAIR mismatch using only DWI imaging.
• To assess the DL model's performance in identifying patients who may benefit from thrombolysis.

Main Methods:

• A retrospective study utilized AIS patient data from the ETIS registry (derivation) and WAKE-UP trial (validation).
• A DL model was trained to predict FLAIR Visible Areas (FVA) using only DWI input.
• The model's predictive value was assessed using Area Under the ROC Curve (AUC) and optimal FVA-index cutoff.

Main Results:

• The DL model demonstrated strong predictive performance for DWI-FLAIR mismatch in both derivation (AUC=0.85) and validation (AUC=0.86) cohorts.
• An optimal FVA-index cutoff of 0.5 achieved 70% sensitivity and 88% specificity in the validation cohort.
• The model showed good agreement (kappa=0.54) with visual assessment of DWI-FLAIR mismatch.

Conclusions:

• The developed DL model accurately predicts DWI-FLAIR mismatch in AIS patients with unknown stroke onset.
• This AI tool can assist clinicians in challenging visual rating scenarios or when FLAIR imaging is unavailable.
• The model holds potential to improve treatment decisions for AIS patients.