Morphology-aware multi-source fusion-based intracranial aneurysms rupture prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Deep learning models for aneurysm rupture prediction typically require large labeled datasets.
• Limited labeled data presents a significant challenge in training accurate predictive models.
• Existing methods may not fully capture complex aneurysm morphology for rupture risk assessment.

Purpose of the Study:

• To develop a deep learning approach for aneurysm rupture prediction using limited labeled data.
• To leverage self-supervised learning to extract meaningful aneurysm morphology embeddings.
• To improve the accuracy of aneurysm rupture prediction compared to existing methods.

Main Methods:

• A self-supervised learning method pretrained a backbone model on unlabeled angiographic images to learn aneurysm morphology embeddings.
• The pretrained model was finetuned using a small set of labeled cases with known rupture status.
• Clinical information was integrated with deep embeddings, and the model was compared against radiomics and conventional morphology models.

Main Results:

• The proposed model achieved an AUC of 0.823, outperforming models trained from scratch (AUC=0.787).
• Integrating clinical information further improved performance to AUC=0.853, significantly outperforming radiomics (AUC=0.805) and conventional morphology models (AUC=0.766).
• An assistive diagnosis system improved neurosurgeons' rupture prediction performance from AUC=0.877 to 0.945.

Conclusions:

• A self-supervised learning approach enables competitive deep learning models for aneurysm rupture prediction with limited data.
• The developed deep embeddings effectively represent aneurysm morphology, leading to superior prediction performance.
• The assistive diagnosis system enhances neurosurgeons' ability to predict aneurysm rupture, offering clinical utility.