Intracranial Aneurysm Segmentation with a Dual-Path Fusion Network

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neurosurgery and neurology

Background:

• Intracranial aneurysms (IAs) are prevalent and life-threatening vascular conditions.
• Accurate diagnosis of IAs is challenged by their small size and variable morphology.
• Automated segmentation of IAs is crucial for diagnostic precision but current deep learning methods often lose detailed information.

Purpose of the Study:

• To introduce an advanced deep learning architecture, the Dual-Path Fusion Network (DPF-Net), for refined intracranial aneurysm segmentation.
• To improve the incorporation of detailed morphological information in IA segmentation models.
• To enhance the accuracy and robustness of automated IA diagnosis.

Main Methods:

• Developed the Dual-Path Fusion Network (DPF-Net) with a resolution-preserving detail branch to minimize information loss.
• Integrated a cross-fusion module to connect semantic and detailed features.
• Employed a detail aggregation module for multi-scale feature fusion and a view fusion strategy to mitigate spatial disruptions.

Main Results:

• DPF-Net achieved a mean Dice Similarity Coefficient (DSC) of 0.8967 on the CADA dataset for IA segmentation.
• Demonstrated robust performance on the BraTS 2020 MRI dataset for brain tumor segmentation, achieving a mean DSC of 0.8535.
• The network effectively preserves and fuses detailed information, enhancing segmentation quality.

Conclusions:

• DPF-Net significantly improves automated segmentation of intracranial aneurysms by integrating detailed information.
• The model's performance highlights its potential for clinical application in automated IA diagnosis.
• DPF-Net shows generalizability and robustness across different medical imaging segmentation tasks.