An automatic machine learning approach for ischemic stroke onset time identification based on DWI and FLAIR imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Thrombolytic therapy for acute ischemic stroke (AIS) is time-sensitive, typically limited to 4.5 hours post-stroke.
• Unknown time since stroke (TSS) frequently excludes patients from this critical treatment.
• Diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) mismatch shows potential for TSS estimation.

Purpose of the Study:

• To develop and validate an automatic machine learning method for classifying TSS relative to the 4.5-hour treatment window.
• To improve patient selection for thrombolytic therapy in AIS by addressing unknown onset times.

Main Methods:

• A cross-modal convolutional neural network was developed for segmenting stroke lesions in DWI and FLAIR images.
• Features were extracted from segmented regions of interest (ROIs) in DWI and FLAIR.
• Machine learning models were trained using these features to classify TSS as <4.5h or >4.5h.

Main Results:

• The method achieved high Dice coefficients for DWI (0.803) and FLAIR (0.647) lesion segmentation.
• The classification model demonstrated an accuracy of 0.805, with 0.769 sensitivity and 0.840 specificity.
• Performance surpassed human reading of DWI-FLAIR mismatch for TSS identification.

Conclusions:

• The proposed automatic machine learning approach effectively classifies time since stroke using DWI and FLAIR imaging.
• This AI-driven method offers a potential solution for fast and automatic TSS identification, expanding treatment eligibility for AIS patients.
• The findings highlight the utility of machine learning in neuroimaging for clinical decision support.