Predicting infarct outcomes after extended time window thrombectomy in large vessel occlusion using knowledge guided...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Neurology

Background:

• Accurate prediction of final infarct volume is crucial for treatment planning in acute ischemic stroke (AIS) patients undergoing mechanical thrombectomy (MT), especially within extended time windows.
• Deep learning models offer potential for improving infarct prediction accuracy using pre-MT brain perfusion data.
• Incorporating prior medical knowledge can enhance the performance of these deep learning models.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting post-mechanical thrombectomy infarct volume in acute ischemic stroke patients.
• To assess the impact of integrating various sources of prior medical knowledge into deep learning models for improved infarct prediction accuracy.
• To compare the performance of different deep learning architectures, including a baseline model and models incorporating collateral flow, infarct probability, and arterial territory mapping.

Main Methods:

• A retrospective study included 221 AIS patients who underwent MT over 6 hours from symptom onset, utilizing pre-MT CT perfusion data.
• Five Swin transformer-based models were developed: BaselineNet, CollateralFlowNet, InfarctProbabilityNet, ArterialTerritoryNet, and UnifiedNet (combining all knowledge sources).
• Model performance was quantified using Dice coefficient and intersection over union (IoU), with post-MT diffusion-weighted imaging serving as ground truth.

Main Results:

• The UnifiedNet model, integrating all prior knowledge, achieved the highest performance with a Dice coefficient of 0.82 and IoU of 0.71.
• Models incorporating medical knowledge significantly outperformed the BaselineNet (Dice 0.69, IoU 0.53).
• Specific knowledge integration showed improvements: CollateralFlowNet (Dice 0.72, IoU 0.56), InfarctProbabilityNet (Dice 0.74, IoU 0.58), and ArterialTerritoryNet (Dice 0.75, IoU 0.60).

Conclusions:

• Integrating medical knowledge into deep learning models substantially enhances the accuracy of infarct predictions in AIS patients undergoing extended time window MT.
• The UnifiedNet model, leveraging multiple knowledge sources, demonstrated superior performance in predicting infarct outcomes.
• These findings support the use of AI-enhanced prediction tools for optimizing treatment strategies in acute ischemic stroke.