High-throughput 3DRA segmentation of brain vasculature and aneurysms using deep learning

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Artificial Intelligence in Medicine

Background:

• Cerebral aneurysm segmentation in 3D rotational angiography (3DRA) is crucial for treatment planning and hemodynamic analysis.
• Challenges include small vasculature volume, class imbalance, and similar appearance of aneurysms and vessels.
• Accurate segmentation is vital for incidental aneurysm detection and risk assessment.

Purpose of the Study:

• To develop a novel multi-class convolutional neural network for precise and robust segmentation of cerebral vessels and aneurysms in 3DRA images.
• To address challenges of class imbalance and inter-class interference in neuroimaging segmentation.
• To facilitate automatic detection and analysis of cerebral vascular structures.

Main Methods:

• A novel multi-class convolutional neural network architecture was proposed.
• The model was trained and evaluated on both internal multi-center and external public datasets.
• Performance was compared against state-of-the-art segmentation approaches.

Main Results:

• The proposed method achieved superior performance in vessel and aneurysm segmentation compared to existing methods.
• Aneurysm segmentation achieved an average Dice score of 0.81 and surface-to-surface error of 0.20 mm.
• Vessel segmentation achieved an average Dice score of 0.91 and surface-to-surface error of 0.25 mm.
• Accurate segmentation was demonstrated in 190 out of 223 clinical cases.

Conclusions:

• The developed approach effectively handles class imbalance and inter-class interference in multi-class segmentation tasks.
• The method shows consistent performance across diverse clinical datasets.
• The segmentation results are suitable for subsequent hemodynamic simulations and pre-operative planning.