Multistage multimodal model predicts 90-day stroke outcomes after thrombectomy

A multimodal deep learning model combining pre- and post-treatment CT/MRI with clinical notes predicted 90-day modified Rankin Scale after stroke thrombectomy with AUC 0.9147 (noncontrast CT) and 0.9479 (MRI), outperforming single-modality baselines.

• The architecture uses a shared foundation model with Low-Rank Adaptation, shared and stage-specific feature extractors, and bidirectional cross-attention between imaging and clinical text.
• Ablation experiments supported the value of multi-stage interaction and multimodal fusion over single-stage or single-modality approaches.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.