Deep learning to identify stroke within 4.5 h using DWI and FLAIR in a prospective multicenter study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Acute ischemic stroke requires timely treatment, with thrombolysis eligibility often limited to within 4.5 hours of symptom onset.
• Determining precise stroke onset can be challenging, particularly in cases like wake-up strokes, impacting treatment decisions.
• Current interpretation of imaging for stroke onset estimation can be variable.

Purpose of the Study:

• To develop and validate a deep learning model for accurate estimation of stroke onset time using multimodal MRI.
• To improve the identification of patients eligible for thrombolysis by classifying stroke onset within or beyond the critical 4.5-hour window.
• To create a robust AI tool that assists clinicians in managing acute ischemic stroke, especially in cases with uncertain symptom onset.

Main Methods:

• Development of a multimodal Res-U-Net (mRUNet) model integrating a modified U-Net and ResNet-34 architecture.
• Utilizing diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) MRI sequences from patients scanned within 24 hours of symptom onset.
• Training and validation of the mRUNet model on internal and two external datasets (single-center and multi-center) with performance comparison against ResNet-34 and DenseNet-121.

Main Results:

• The mRUNet model achieved high performance in classifying stroke onset within the 4.5-hour window, with AUC-ROC values of 0.903 (internal), 0.910 (external single-center), and 0.868 (external multi-center).
• mRUNet demonstrated significantly superior performance compared to ResNet-34 and DenseNet-121 across all evaluated datasets.
• The model showed robust and consistent classification capabilities for the 4.5-hour onset-time window across diverse patient cohorts.

Conclusions:

• The developed mRUNet deep learning model accurately and reliably estimates acute ischemic stroke onset using DWI and FLAIR MRI.
• This AI tool has the potential to support timely and individualized thrombolysis decisions in clinical practice, particularly for patients with unclear stroke onset.
• Leveraging multimodal MRI data, mRUNet acts as a 'tissue clock' to enhance acute stroke management and improve patient outcomes.