Deep geometric learning for intracranial aneurysm detection: towards expert rater performance
Žiga Bizjak1, June Ho Choi2, Wonhyoung Park2
1Laboratory of Imaging Technologies, University of Ljubljana Faculty of Electrical Engineering, Ljubljana, Slovenia ziga.bizjak@fe.uni-lj.si.
Journal of Neurointerventional Surgery
|October 13, 2023
Summary
Deep learning accurately detects intracranial aneurysms (IAs) on MRA and CTA scans, matching expert sensitivity, especially for small IAs. This AI approach offers a promising tool for computer-assisted IA detection in clinical settings.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurosurgery
Background:
- Early detection of intracranial aneurysms (IAs) is critical for improving patient outcomes.
- Current methods like CT angiography (CTA) and MR angiography (MRA) have moderate sensitivity for small IAs (<3mm).
- Deep learning shows potential to achieve expert-level sensitivity in IA detection.
Purpose of the Study:
- To develop and validate a novel deep learning approach for detecting intracranial aneurysms.
- To achieve expert-level sensitivity in IA detection across different imaging modalities.
- To assess the performance of the AI model in a clinical setting.
Main Methods:
- A large, multisite dataset of 1054 MRA and 2174 CTA scans with expert annotations was utilized.
- A two-step, modality-agnostic deep learning approach was developed.
- The method involved vascular segmentation using nnU-Net and aneurysm classification using PointNet++ on sampled point clouds.
Main Results:
- The approach achieved pooled sensitivities of 85% (MRA) and 90% (CTA) on external validation data.
- Sensitivity for small IAs (<3mm) was 72% (MRA) and 83% (CTA).
- Low false finding rates were observed (1.54-1.57 per image) and minimal on healthy data (0.4-0.83).
Conclusions:
- The proposed deep learning method achieves state-of-the-art performance in IA detection.
- It matches expert-level sensitivity for both small and larger IAs.
- The approach is suitable for clinical implementation as a computer-assisted detection tool for IAs.
