Differential Diagnostic Value of Machine Learning-Based Models for Embolic Stroke.

Area of Science:

• Neurology
• Oncology
• Radiology
• Artificial Intelligence

Background:

• Cancer-associated thrombosis (CAT) and atrial fibrillation (AF)-related stroke are distinct embolic stroke subtypes.
• These subtypes exhibit unique diffusion-weighted imaging (DWI) lesion patterns.
• Accurate differentiation is crucial for appropriate patient management.

Purpose of the Study:

• To assess the feasibility of using DWI-based machine learning models for differentiating CAT from AF-related stroke.
• To evaluate the performance of convolutional neural network (CNN) models in this diagnostic task.
• To explore the impact of data augmentation on model performance.

Main Methods:

• A pilot study enrolled patients with CAT and AF-related stroke.
• Diffusion-weighted imaging (DWI) data were augmented using flipping and contrast shifting.
• Convolutional neural network (CNN) models were developed and trained using DWI images, with one model also incorporating demographic/clinical data.

Main Results:

• The DWI-based CNN model achieved training accuracy of 87.1% and testing accuracy of 78.6%.
• The CNN model combining DWI images with clinical data achieved higher accuracies (training: 95.2%, testing: 85.7%).
• No significant differences were observed in sensitivity, specificity, accuracy, or AUC between the two models.

Conclusions:

• DWI-based CNN models, especially with data augmentation, show potential for differentiating CAT from AF-related stroke.
• Combining imaging data with clinical information may enhance diagnostic performance.
• Further validation in larger cohorts is warranted.