Detection of Onset Time for Acute Ischemic Stroke Based on Multi-scale Features and Cross-Attention

Acute ischemic stroke (AIS) is a cerebral disease that can lead to severe brain tissue damage and even death. The optimal treatment window for AIS is within 6 h since time since stroke (TSS). Computed tomography perfusion (CTP) provides crucial brain-related information and plays a significant role in the diagnosis and treatment of AIS. However, CTP data poses challenges due to its small sample size, high dimensionality, and the heterogeneity and interdependence among the four types of perfusion images. To effectively utilize CTP data for accurately assessing the time window and preventing disease progression, we propose a classification model based on multi-scale feature fusion and cross-attention mechanisms tailored to the clinical characteristics of CTP. Specifically, we employ a multi-scale feature extraction (MFE) blocks to combine features from various scales and utilize a cross-attention fusion (CAF) module to merge complementary features. Finally, the multi-head pooling attention (MPA) is employed to further learn feature information and obtain as much critical information as possible. To validate the effectiveness of our proposed approach on a hospital's private dataset, we conduct a 5-fold cross-validation strategy. Experimental results demonstrate that our method exhibits superiority in the TSS classification task and possesses high robustness.