A cross-attention-based deep learning approach for predicting functional stroke outcomes using 4D CTP imaging and...

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence

Background:

• Acute ischemic stroke (AIS) poses a significant global health burden, often resulting in long-term disability.
• Spatio-temporal (4D) Computed Tomography Perfusion (CTP) is vital for assessing brain tissue in AIS.
• Predicting functional outcomes using 4D CTP and clinical data remains an underexplored area.

Purpose of the Study:

• To develop and evaluate a novel multimodal deep learning model for predicting 90-day modified Rankin Scale in AIS patients.
• To integrate 4D CTP imaging with clinical metadata for enhanced outcome prediction.
• To investigate the efficacy of intermediate fusion with cross-attention for combining multimodal data.

Main Methods:

• A multimodal deep learning model was developed, employing an intermediate fusion strategy with cross-attention.
• The model combined features from 4D CTP scans and patient clinical metadata.
• The model was evaluated on a dataset of 70 AIS patients who underwent endovascular mechanical thrombectomy.

Main Results:

• The proposed multimodal model achieved an accuracy of 0.77 in predicting functional outcomes.
• This outperformed late fusion strategies (0.73 accuracy) and unimodal models (0.61 for 4D CTP, 0.71 for clinical data).
• The cross-attention mechanism effectively leveraged inter-modal relationships.

Conclusions:

• Multimodal deep learning, particularly with intermediate fusion, significantly improves the prediction of functional outcomes in AIS.
• Combining 4D CTP and clinical data offers a more robust approach than using either modality alone.
• Advanced fusion techniques hold promise for personalized stroke management and treatment planning.