3D Intracranial Aneurysm Classification and Segmentation via Unsupervised Dual-Branch Learning

Area of Science:

• Medical Imaging and Artificial Intelligence
• Digital Health and Neuroimaging

Background:

• Intracranial aneurysms are a significant concern in digital health.
• Current deep learning methods for aneurysm detection primarily rely on supervised approaches using medical images.

Purpose of the Study:

• To develop an unsupervised deep learning method for detecting intracranial aneurysms.
• To utilize 3D point cloud data for aneurysm detection, overcoming limitations of supervised methods.

Main Methods:

• The proposed method involves two stages: unsupervised pre-training and downstream tasks.
• Unsupervised pre-training utilizes a dual-branch contrastive network to maximize correspondence between original and jittered point clouds.
• Downstream tasks employ simple networks for supervised classification and segmentation.

Main Results:

• The unsupervised method demonstrated comparable or superior performance to state-of-the-art supervised techniques on the IntrA dataset.
• The method showed particular prominence in the detection of aneurysmal vessels.
• Achieved 90.79% accuracy on the ModelNet-40 dataset, outperforming existing unsupervised models.

Conclusions:

• Unsupervised learning on 3D point cloud data offers a viable and effective alternative for intracranial aneurysm detection.
• The developed method shows promise for advancing intelligent detection in digital health applications.
• The approach is robust and achieves high accuracy, even outperforming supervised methods in specific tasks.