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

Abstract

Identifying large vessel occlusion (LVO) is of significant importance for the treatment and prognosis of acute ischemic stroke (AIS) patients. CT Angiography (CTA) is commonly used in LVO identification due to its visibility of vessels and short acquisition time. It is challenging to make LVO identification methods focus on vascular regions without vessel segmentation while accurate vessel segmentation is difficult and takes more time. Meanwhile, most existing methods fail to effectively integrate clinical prior knowledge. In this work, we propose VANet, a novel LVO identification network which utilizes coarse-grained vessel feature for feature enhancement and learns asymmetry of two brain hemispheres on CTA of AIS patients. Firstly, we reconstruct 3D CTA scans into 2D based on maximum intensity projection (MIP) to reduce computational complexity and highlight vessel information. Secondly, we design a coarse-grained vessel aware module based on simple edge detection and morphological operations to acquire coarse-grained vessel feature without precise vessel segmentation. Thirdly, we design a vessel-guided feature enhancement that directs the model's attention to vessel areas in the images by utilizing coarse-grained vessel feature. Finally, inspired by the clinical knowledge that LVO can lead to asymmetry in brain, we design an asymmetry learning module utilizing deep asymmetry supervision to keep the patients' inherent asymmetry invariant and using asymmetry computing to acquire effective asymmetry features. We validate the proposed VANet on our private internal and external AIS-LVO datasets which contain 366 and 81 AIS patients, respectively. The results indicate that our proposed VANet achieves an accuracy of 94.54% and an AUC of 0.9685 on the internal dataset, outperforms 11 state-of-the-art methods (including general classification methods and LVO-specific methods). Besides, our method also achieves the best accuracy of 88.89% and AUC of 0.9111 when compared to 11 methods on the external test dataset, implying its good generalization ability. Interpretability analysis shows that the proposed VANet can effectively focus on vascular regions and learn asymmetry features.