3D foundation model for generalizable disease detection in head computed tomography.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Head computed tomography (CT) is crucial for diagnosing brain, skull, and cerebrovascular conditions.
• Its rapid acquisition and availability make it ideal for neurologic emergencies.
• Developing deep learning models for disease detection is limited by scarce, high-quality annotations.

Purpose of the Study:

• To introduce FM-HCT, a Foundation Model for Head CT.
• To enable generalizable disease detection using self-supervised learning.
• To overcome annotation scarcity in developing powerful deep learning models for head CT.

Main Methods:

• Pretraining a deep learning model on 361,663 non-contrast 3D head CT scans.
• Utilizing self-supervised learning, eliminating the need for manual annotations.
• Developing a foundation model (FM-HCT) to learn robust, generalizable features.

Main Results:

• The self-supervised foundation model learned robust and generalizable features.
• FM-HCT demonstrated improved performance on downstream diagnostic tasks.
• Performance gains were significant compared to models trained from scratch and previous 3D CT foundation models.

Conclusions:

• Self-supervised learning with foundation models offers a powerful approach for head CT analysis.
• FM-HCT addresses the challenge of limited annotations in medical imaging AI.
• This method enhances the generalizability and diagnostic capabilities of deep learning models for head CT.