Deep learning-based recognition and segmentation of intracranial aneurysms under small sample size

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Manual segmentation of intracranial aneurysms (IAs) for 3D reconstruction is time-consuming and error-prone.
• Accurate and rapid patient-specific IA reconstruction is crucial for clinical management and research.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated IA identification and segmentation.
• To compare the performance of different convolutional neural network (CNN) architectures for IA segmentation.

Main Methods:

• A deep learning framework utilizing 3D segmentation-dedicated CNNs (3D UNet, VNet, 3D Res-UNet) was developed.
• The framework was trained and evaluated on 101 cranial computed tomography angiography (CTA) datasets with 140 IA cases.
• Performance was assessed using metrics including Voxel-wise Recall (V-Recall), Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and segmentation time.

Main Results:

• 3D UNet demonstrated superior segmentation performance, achieving an average V-Recall of 0.797 and DSC of 0.818.
• 3D UNet showed a lower average Hausdorff Distance (3.323 voxels) and minimal deviation in 3D analysis.
• The average segmentation time for 3D UNet was efficient at 0.053s.

Conclusions:

• The proposed deep learning framework, particularly with 3D UNet, provides a fast and accurate solution for IA identification and segmentation.
• This automated approach can significantly benefit routine clinical management and large-scale cohort studies of IAs.
• Deep learning models show promise in overcoming the limitations of manual segmentation in neuroimaging analysis.