Predicting final ischemic stroke lesions from initial diffusion-weighted images using a deep neural network

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Neurology

Background:

• Diffusion-perfusion mismatch is crucial for estimating stroke infarct risk without reperfusion.
• Perfusion-weighted imaging (PWI) prolongs and increases the cost of acute stroke workups.
• Exploring AI to predict stroke outcomes using limited imaging data is essential.

Purpose of the Study:

• To determine if a deep convolutional neural network (DCNN) trained on diffusion-weighted imaging (DWI) can predict final infarct volume and location.
• To assess the feasibility of reducing acute stroke imaging time and cost.
• To develop a predictive model for stroke prognosis using readily available admission imaging.

Main Methods:

• An attention-gated DCNN (AG-DCNN) was trained and validated on 445 acute stroke patients.
• Input data included DWI, apparent diffusion coefficient (ADC) maps, and thresholded ADC maps.
• The model predicted voxel-wise infarction probability maps, with performance evaluated using AUC, DSC, and volume error.

Main Results:

• The AG-DCNN achieved a median AUC of 0.91.
• Median sensitivity and specificity for infarction prediction were 0.60 and 0.97, respectively.
• The model demonstrated good correlation (ρc = 0.73) between predicted and true infarct volumes.

Conclusions:

• An AG-DCNN utilizing only admission DWI can predict 3-7 day infarct volumes accurately.
• This AI model shows comparable accuracy to methods using both DWI and PWI.
• The findings suggest potential for shorter stroke imaging protocols, facilitating quicker treatment decisions.