Dual-attention and multi-task learning for onset time assessment in acute ischemic stroke

Background: Accurately determining the onset time of acute ischemic stroke (AIS) is crucial for guiding reperfusion therapy. Traditional methods such as diffusion-weighted imaging (DWI)-fluid-attenuated inversion recovery (FLAIR) mismatch are limited by high subjectivity and inconsistency. Methods: This study collected DWI and FLAIR sequences from AIS patients with a confirmed onset time within 24 h. We proposed a lightweight Dual Attention-enhanced Multi-task Network (DAMN), which integrates multi-sequence fusion and multi-task learning. The model’s performance in classifying onset time across three critical windows (4.5 h, 6 h, 9 h) was evaluated against baseline models and single-sequence inputs. Ablation studies validated the contribution of the dual-attention mechanism. Results: From 218 included patients, 8744 aligned axial slices were obtained. For lesion identification, the multi-sequence DAMN achieved an area under the ROC curve (AUC) of 0.972. For onset time classification, it consistently outperformed single-sequence models across all time windows, with AUCs of 0.827 (4.5 h), 0.820 (6 h), and 0.810 (9 h). In the 4.5-hour task, the dual-attention mechanism provided a significant performance gain of 0.118 in AUC over the baseline backbone. Conclusions: The proposed DAMN model, leveraging multi-sequence MRI fusion and dual-attention, effectively focuses on lesion areas and demonstrates superior performance for the critical 4.5-hour onset time classification, offering valuable decision support for thrombolytic therapy. The source code is publicly available at: https://github.com/813Ender/Wus-code.git .