Deep learning-based functional outcome prediction of acute ischemic stroke using DWI, SWI imaging and clinical metadata

Shannan Chen¹Hongyi Li²Xuanhe Zhao¹Lingkai Liu¹Yang Duan³Wei Qian¹Peizhuo Zang²Chao Li⁴Ronghui Ju²Shouliang Qi^1*

• ¹Northeastern University, Shenyang, China
• ²People's Hospital of Liaoning Province, Shenyang, China
• ³General Hospital of Northern Theatre Command, Shenyang, China
• ⁴University of Cambridge Department of Medicine, Cambridge, United Kingdom

Predicting functional outcomes in acute ischemic stroke (AIS) is crucial for personalized treatment yet often relies on time-consuming manual annotation. Our study introduces a deep learning framework that automatically segments stroke lesions and deep medullary veins (DMVs) on diffusion-weighted imaging (DWI) and susceptibility-weighted imaging (SWI), respectively, and predicts 90-day modified Rankin Scale outcomes by integrating multimodal imaging with clinical metadata. We included 150 AIS patients for model development and 20 for external validation. Our framework integrates DWI, SWI, and clinical metadata. A diffusion model-based segmentation network identifies stroke lesions and DMVs, while a Generative Adversarial Network based data augmentation method generates diverse lesion and vein structures to boost segmentation generalization. Finally, a multimodal classification model fuses segmented images and clinical metadata via a random forest classifier and Dempster-Shafer theory for robust functional outcome prediction. After data augmentation, the segmentation performance reached 80.05% Dice score (DWI) and 40.66% Dice score (SWI), with accurate boundary delineation (HD95: 8.14 mm and 10.59 mm). Regarding functional outcome prediction, the proposed method achieved 93.3% accuracy on the in-house dataset and 80.0% on the external dataset, respectively. Our proposed method outperformed existing methods in accuracy and consistency in both lesion segmentation and functional outcome prediction for AIS. This study presents an automated framework to predict AIS functional outcomes by synergizing DWI, SWI and clinical metadata. Its high accuracy and external validation address the limitations of manual assessment, offering a scalable tool for optimizing stroke management and supporting clinicians in treatment decisions.