Generalizable self-supervised learning for brain CTA in acute stroke.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Medical Imaging
• Computational Neuroscience

Background:

• Acute stroke management relies on timely interpretation of CT angiography (CTA) data.
• Current models lack generalizability across multiple acute stroke tasks and struggle with unlabeled data.
• 3D feature representation is crucial for complex tasks like large vessel occlusion detection.

Purpose of the Study:

• To develop generalizable models for multiple acute stroke tasks using unlabeled data.
• To leverage self-supervised contrastive learning with 3D CTA images and radiology report findings for pretraining.
• To evaluate the impact of linear probing on model generalizability and performance.

Main Methods:

• Utilized a self-supervised contrastive learning framework for pretraining on 1,542 pairs of 3D CTA scans and radiology reports.
• Applied the pretrained model to four tasks: large vessel occlusion detection, acute ischemic stroke detection, intracerebral hemorrhage classification, and ischemic core volume prediction.
• Evaluated model performance through fine-tuning and testing on a separate dataset of 592 subjects, assessing the influence of linear probing.

Main Results:

• Linear probing during pretraining significantly enhanced model generalizability and classification performance.
• The best-performing models demonstrated robust generalization to out-of-distribution data across multiple tasks.
• Pretraining with radiology report findings provided substantial performance gains compared to training solely on labeled data.

Conclusions:

• Self-supervised contrastive learning with 3D CTA and report findings offers a powerful approach for acute stroke task generalizability.
• Linear probing during pretraining improves predictive performance and robustness.
• This method advances AI-driven acute stroke diagnosis and management, particularly for challenging tasks.