A self-supervised learning approach for registration agnostic imaging models with 3D brain CTA.

Area of Science:

• Neuroimaging
• Deep Learning
• Medical Diagnostics

Background:

• Deep learning models for acute stroke neuroimaging often require image registration, adding computational cost and potential failure points.
• Adapting existing models to handle unregistered images is challenging but crucial for broader applicability.

Purpose of the Study:

• To develop a general-purpose contrastive self-supervised learning method for adapting deep neural networks to process unregistered neuroimaging data.
• To enable convolutional neural networks trained on registered images to function effectively on a different input domain (unregistered images).

Main Methods:

• A contrastive self-supervised learning strategy was employed, utilizing a teacher-student network architecture without requiring manual labels.
• The original model (teacher) guided the training of a new network (student) to learn from unregistered images.

Main Results:

• The student model, trained on unregistered computed tomographic angiography (CTA) data, achieved competitive large vessel occlusion (LVO) detection performance (AUC = 0.88) compared to the teacher model (AUC = 0.81).
• Directly training a student model on unregistered images using standard supervised learning resulted in significantly lower performance (AUC = 0.63).

Conclusions:

• The proposed self-supervised method effectively adapts deep learning models for neuroimaging analysis, even when dealing with unregistered images.
• This approach demonstrates a viable strategy for improving the robustness and applicability of AI in acute stroke diagnostics without relying on image registration.