Large vessel occlusion identification network with vessel guidance and asymmetry learning on CT angiography of acute ischemic stroke patients

Hulin Kuang ¹, Xinyuan Liu ¹, Jin Liu ¹

⁰Hunan Provincial Key Lab on Bioinformatics, School of Computer Science and Engineering, Central South University, Changsha 410083, China.

Medical Image Analysis +

|

February 11, 2025

View abstract on PubMed

Summary

VANet accurately identifies large vessel occlusion (LVO) in acute ischemic stroke (AIS) patients using CT Angiography (CTA) by learning vascular features and brain asymmetry. This novel network achieves high accuracy and outperforms existing methods, improving stroke diagnosis.

Area Of Science

Medical Imaging
Artificial Intelligence
Neurology

Background

Accurate identification of large vessel occlusion (LVO) is critical for acute ischemic stroke (AIS) patient treatment and prognosis.
CT Angiography (CTA) is a common imaging modality for LVO detection due to its speed and vessel visualization.
Existing LVO identification methods struggle with precise vessel segmentation and integrating clinical knowledge, often requiring time-consuming segmentation processes.

Purpose Of The Study

To propose VANet, a novel network for LVO identification in AIS patients using CTA.
To develop a method that focuses on vascular regions without requiring precise vessel segmentation.
To integrate clinical prior knowledge, specifically brain asymmetry, into the LVO identification process.

Main Methods

Reconstruction of 3D CTA scans into 2D using Maximum Intensity Projection (MIP) to simplify computation and enhance vessel visibility.
Implementation of a coarse-grained vessel aware module using edge detection and morphological operations for feature extraction without segmentation.
Development of a vessel-guided feature enhancement module and an asymmetry learning module incorporating deep supervision and asymmetry computing to leverage clinical insights.

Main Results

VANet achieved 94.54% accuracy and 0.9685 AUC on an internal AIS-LVO dataset (366 patients), outperforming 11 state-of-the-art methods.
On an external dataset (81 patients), VANet demonstrated strong generalization with 88.89% accuracy and 0.9111 AUC, also outperforming 11 comparative methods.
Interpretability analysis confirmed VANet's ability to focus on relevant vascular regions and effectively learn asymmetry features.

Conclusions

The proposed VANet effectively identifies LVO in AIS patients by utilizing coarse-grained vessel features and learning hemispheric asymmetry from CTA.
VANet offers a promising alternative to methods requiring precise vessel segmentation, improving efficiency and accuracy in stroke diagnosis.
The network's strong performance on both internal and external datasets highlights its robustness and potential for clinical application in acute stroke management.

Keywords:

Acute ischemic stroke Brain asymmetry learning CT angiography Coarse-grained vessel awareness Large vessel occlusion identification