Ischemic stroke subtyping method combining convolutional neural network and radiomics.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurology

Background:

• Cardiogenic embolism (CE) and large-artery atherosclerosis embolism (LAA) are the most prevalent ischemic stroke (IS) subtypes.
• Accurate subtyping of IS is crucial for effective patient diagnosis and treatment.

Purpose of the Study:

• To develop and validate an IS subtyping method integrating convolutional neural networks (CNN) and radiomics.
• To enhance the precision of IS diagnosis and assist clinicians in treatment planning.

Main Methods:

• Segmentation of brain embolism regions from computed tomography angiography (CTA) images and extraction of radiomics features.
• Optimization of radiomics features using L2 norm and feature selection via random forest.
• Extraction of deep learning features using a CNN Cap-UNet model.
• Training and selection of optimal IS subtyping classification models by combining radiomics and deep learning features.

Main Results:

• The study utilized CTA images from 82 IS patients.
• The optimal subtyping model, employing an Adaboost classifier, achieved an AUC of 0.9018 and an accuracy of 0.8929.
• The proposed method demonstrated high performance in IS subtyping.

Conclusions:

• The developed method effectively predicts IS subtypes.
• This approach shows significant potential to support clinicians in making timely and accurate diagnoses for IS patients.