Both Infarcted and Noninfarcted Brain Regions Contribute to Deep Learning-Based MRI Prediction of Acute Stroke...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Stroke Medicine

Background:

• Predicting long-term clinical outcomes after acute ischemic stroke (AIS) is crucial for patient care and research.
• Early information from brain imaging can inform these predictions.

Purpose of the Study:

• To investigate the contribution of different brain imaging regions, including non-infarcted areas, to predicting 90-day stroke outcomes using deep learning (DL).
• To assess the efficacy of DL models trained on various imaging inputs for stroke outcome prediction.

Main Methods:

• Developed and validated DL models on MRI DWI scans from 449 AIS patients (1-7 days post-stroke).
• Trained models on infarct volumes, full-brain images, infarct masks, intensity-preserved infarct masks, and lesion-neutralized images.
• Evaluated performance using accuracy, mean absolute error (MAE), and area under the curve (AUC) for predicting modified Rankin Scale (mRS) scores.

Main Results:

• The model using full-brain images achieved the lowest MAE (1.07) and highest AUC (0.86) for predicting unfavorable outcomes (mRS > 2).
• Models trained on infarct volume alone showed the highest MAE (1.51).
• Intermediate inputs (masks, lesion-neutralized images) improved predictions over volume alone but underperformed full-brain models.

Conclusions:

• Full brain imaging provides the most accurate predictions for 90-day stroke outcomes.
• Infarct characteristics and non-infarcted brain regions significantly contribute to outcome prediction.
• Non-infarcted areas may reflect brain health and resilience, offering valuable prognostic information.