Multi-Scope Feature Extraction for Intracranial Aneurysm 3D Point Cloud Completion.

Area of Science:

• Medical Imaging
• Computer Vision
• Neuroscience

Background:

• 3D point clouds are increasingly used in medicine, but their application to intracranial vessels and aneurysms is limited by complex reconstruction.
• Accurate 3D representation of intracranial vessels and aneurysms is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To propose a novel deep learning method, the Multi-Scope Feature Extraction Network (MSENet), for 3D point cloud completion of intracranial aneurysms.
• To address the challenge of incomplete data acquisition from different angles in medical scans.

Main Methods:

• Simulated incomplete 3D point clouds of intracranial vessels and aneurysms.
• Developed MSENet with a multi-scope feature extraction encoder for global feature extraction and a folding-based decoder for shape reconstruction.
• Employed residual linking with original point coordinates for geometric accuracy and refined the output through merging and sampling.

Main Results:

• MSENet significantly improved completion accuracy on 3D intracranial aneurysm datasets and general 3D vision datasets.
• Achieved a 21.1% increase in F-score and a 33.8% reduction in Chamfer Distance compared to baseline methods.
• Demonstrated superior performance across multiple evaluation metrics, including Earth Mover's Distance.

Conclusions:

• MSENet offers an effective solution for 3D point cloud completion of intracranial aneurysms, overcoming limitations of traditional methods.
• The proposed network enhances the potential of 3D point cloud data in neurovascular imaging and analysis.