Deep learning-based fine-grained assessment of aneurysm wall characteristics using 4D-CT angiography

Area of Science:

• Medical imaging analysis
• Computational fluid dynamics
• Machine learning in healthcare

Background:

• Cerebral aneurysms pose significant health risks.
• Differentiating aneurysm wall characteristics is crucial for risk stratification.
• Current methods for evaluating aneurysm wall properties are limited.

Purpose of the Study:

• To develop a novel deep learning approach for characterizing aneurysm wall regions (thin-walled and hyperplastic-remodeling).
• To evaluate aneurysm wall characteristics using 4D-computed tomography angiography (4D-CTA) data.
• To enable point-level risk evaluation through time-series analysis of aneurysm wall motion.

Main Methods:

• Analysis of 52 unruptured cerebral aneurysms using 4D-CTA and intraoperative recordings.
• Development of a CNN-LSTM regression model with attention layers for time-series data.
• Implementation of patient-independent feature extraction and incorporation of unlabeled data.

Main Results:

• The deep learning model achieved an average diagnostic accuracy of 92%.
• The attention-based model significantly outperformed a simpler model without attention.
• Patient-independent feature extraction and unlabeled data enhanced model performance.

Conclusions:

• Deep learning effectively predicts aneurysm wall characteristics from 4D-CTA.
• Attention mechanisms in deep learning models are critical for enhanced performance.
• The proposed method offers a promising tool for fine-grained aneurysm risk assessment.