Diffusion-/perfusion-weighted imaging fusion to automatically identify stroke within 4.5 h.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate and timely stroke identification is critical for effective treatment.
• Machine learning (ML) models offer potential for improving diagnostic accuracy in acute stroke.
• Integrating diffusion-weighted imaging (DWI) and perfusion-weighted imaging (PWI) may enhance stroke detection.

Purpose of the Study:

• To develop and compare ML models based on diffusion- and perfusion-weighted imaging fusion (DP fusion) for identifying stroke within 4.5 hours.
• To evaluate DP fusion ML models against DWI- and/or PWI-based ML models.
• To create and assess an automatic segmentation-classification model (X-Net) for stroke identification.

Main Methods:

• Developed ML models using multimodal MRI data from acute stroke patients.
• Employed manual segmentation, registration, DP fusion, and feature extraction.
• Established logistic regression (LR) and support vector machine (SVM) models, alongside the proposed X-Net segmentation-classification model.
• Evaluated performance using AUC, sensitivity, Dice coefficients, decision curve analysis, and calibration curves.

Main Results:

• DP fusion models achieved the highest AUC for identifying stroke onset time (LR: 0.91, SVM: 0.90 in test sets).
• The DP fusion-LR model demonstrated superior net benefits across risk thresholds with good calibration (average absolute error: 0.049).
• The X-Net model achieved high Dice coefficients (DWI: 0.81, Tmax: 0.83) and performance similar to manual labeling (AUC: 0.84).

Conclusions:

• DP fusion-based ML models are highly effective for identifying stroke within 4.5 hours.
• The automatic X-Net model demonstrates robust performance in segmenting acute stroke lesions.
• These advanced imaging fusion and ML techniques can aid clinical decision-making for acute stroke patients, particularly those with unknown onset times.