Predicting Hypoperfusion Lesion and Target Mismatch in Stroke from Diffusion-weighted MRI Using Deep Learning.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Perfusion imaging is crucial for identifying target mismatch in stroke but requires contrast agents and complex postprocessing.
• Accurate identification of hypoperfusion lesions and target mismatch is essential for effective stroke treatment.
• Current methods for assessing stroke severity and mismatch can be resource-intensive and time-consuming.

Purpose of the Study:

• To develop and validate a deep learning model for predicting stroke hypoperfusion lesions.
• To identify patients with a target mismatch profile using only diffusion-weighted imaging (DWI) and clinical information.
• To compare the deep learning model's performance against established clinical-DWI mismatch criteria.

Main Methods:

• A three-dimensional U-Net deep learning model was trained on retrospective multicenter data from acute ischemic stroke patients.
• Inputs included baseline DWI, apparent diffusion coefficient (ADC) maps, NIH Stroke Scale score, and stroke symptom sidedness.
• Model performance was evaluated using Dice score coefficient (DSC) and compared to the DAWN trial criteria via McNemar test.

Main Results:

• The deep learning model achieved a median DSC of 0.61 for predicting hypoperfusion lesions.
• The model identified target mismatch with 90% sensitivity and 77% specificity.
• This significantly outperformed the clinical-DWI mismatch approach (50% sensitivity, 89% specificity).

Conclusions:

• A 3D U-Net deep learning model can effectively predict hypoperfusion lesions from DWI and clinical data.
• The model demonstrates superior sensitivity in identifying patients with a target mismatch profile compared to existing methods.
• This AI-driven approach holds promise for streamlining stroke assessment and improving patient selection for reperfusion therapies.