Ischemic stroke outcome prediction with diversity features from whole brain tissue using deep learning network.

Area of Science:

• Neuroimaging
• Radiomics
• Machine Learning

Background:

• Ischemic stroke outcome prediction is crucial for effective treatment.
• Current methods often rely on lesion-specific information, which can be time-consuming to analyze.
• There is a need for faster, more accurate prediction methods applicable to the entire brain.

Purpose of the Study:

• To develop and validate a novel outcome prediction method for ischemic stroke.
• To improve the accuracy and efficiency of stroke outcome prediction using whole brain features.
• To assess the utility of combined radiomics and encoding features in predicting stroke prognosis.

Main Methods:

• Extracted dynamic radiomics features (DRFs) from dynamic susceptibility contrast perfusion-weighted imaging (DSC-PWI).
• Extracted static radiomics features (SRFs) and static encoding features (SEFs) from DSC-PWI derived minimum intensity projection (MinIP) maps.
• Utilized Lasso algorithm for feature selection to derive fused features (F_fuse) and evaluated using machine and deep learning models.

Main Results:

• The fused features (F_fuse) derived from DRFs, SRFs, and SEFs, particularly with the Resnet 18 model, demonstrated superior performance.
• Achieved a high mean score of 0.971 in predicting stroke outcomes across both machine learning and deep learning models.
• Deep learning models generally outperformed traditional machine learning models in this prediction task.

Conclusions:

• The proposed method accurately predicts stroke outcomes without the need for ischemic lesion segmentation.
• This approach offers significant clinical value for rapid, efficient, and precise stroke treatment.
• Whole brain feature analysis provides a robust alternative for ischemic stroke outcome prediction.