# An Automatic Detection Method Of Cerebral Aneurysms In Time-Of-Flight Magnetic Resonance Angiography Images Based On Attention 3D U-Net

**Authors:** Chen Geng, Meng Chen, Ruoyu Di, Dongdong Wang, Liqin Yang, Wei Xia, Yuxin Li, Daoying Geng  
**arXiv:** [2110.13367](https://arxiv.org/abs/2110.13367)  
**Date:** 2021-10-26  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2110.13367v1/

## Key points

- Subarachnoid hemorrhage caused by ruptured cerebral aneurysm can be fatal
- Early detection and treatment during asymptomatic periods greatly reduce rupture risk
- Time-of-flight magnetic resonance angiography commonly used for screening cerebral aneurysms
- Recent advancements in deep learning technology show promise in improving detection accuracy
- Researchers proposed a novel method for detecting aneurysms using a fully automatic cerebral artery segmentation algorithm and 3D U-Net model with 3D SENet module
- Study involved 231 patients with unruptured cystic aneurysms undergoing contrast unenhanced 3D TOF-MRA imaging
- Proposed method achieved high sensitivity rates: 97.89% sensitivity in five-fold cross-validation, 91.0% sensitivity with only 2.48 false positives per case in external test sets
- Method demonstrated competitive sensitivity with less training data and low false positive rate compared to previous studies
- Study showcased feasibility and superior performance of deep learning networks in enhancing cerebral aneurysm detection accuracy

## Summary

Background: Subarachnoid hemorrhage caused by ruptured cerebral aneurysm can have fatal consequences. However, early detection and treatment during asymptomatic periods can greatly reduce the risk of rupture. Time-of-flight magnetic resonance angiography is a non-invasive screening technique commonly used for cerebral aneurysms. Recent advancements in deep learning technology have shown promise in improving detection accuracy. Previous studies have emphasized the importance of three-dimensional features in aneurysm detection but have faced challenges such as high false positive rates and the need for extensive training data. In this study, researchers proposed a novel method for detecting aneurysms. They utilized a fully automatic cerebral artery segmentation algorithm to extract the volume of interest without requiring training data. The 3D U-Net model was then improved by incorporating a 3D SENet module to enhance aneurysm detection capabilities. This resulted in a set of fully automated end-to-end methods for detecting aneurysms. The study involved 231 patients with unruptured cystic aneurysms who underwent contrast unenhanced 3D TOF-MRA imaging. The dataset included training sets, internal test sets, and external test sets. By comparing different attention modules added to the original network and optimizing the best solution while increasing the dataset size, researchers were able to improve method performance. Results showed that the proposed method achieved high sensitivity rates with 97.89% sensitivity in five-fold cross-validation and 91.0% sensitivity with only 2.48 false positives per case in external test sets. Compared to previous studies, this method demonstrated competitive sensitivity with less training data while maintaining a low false positive rate. By utilizing 3D U-Net for aneurysm detection and exploring the potential of channel attention mechanisms, this study showcased the feasibility and superior performance of deep learning networks in enhancing cerebral aneurysm detection accuracy. Overall, this research contributes valuable insights into improving screening techniques for cerebral aneurysms through advanced imaging technologies and deep learning algorithms, ultimately offering potential benefits in early diagnosis and treatment planning for at-risk patients.