Multimodal CT Perfusion-Based Deep Learning for Predicting Stroke Lesion Outcomes in Complete and No Recanalization...

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate prediction of final lesion location and volume in acute ischemic stroke (AIS) is critical for effective clinical management.
• Current CT perfusion (CTP) imaging methods for estimating lesion outcomes have limitations.
• Deep learning models offer a promising alternative for improved outcome prediction in AIS.

Purpose of the Study:

• To develop and evaluate specialized deep learning models for predicting final infarct lesions in AIS patients.
• To predict infarct core in cases of successful reperfusion (complete recanalization, CR).
• To predict the combined core-penumbra region in cases of unsuccessful reperfusion (no recanalization, NR).

Main Methods:

• Developed single-modal and multi-modal deep learning models using CTP parameter maps.
• Trained and evaluated models on a multi-center dataset, separating patients into CR (n=350) and NR (n=138) groups.
• Utilized five-fold cross-validation for robust model evaluation.

Main Results:

• The multi-modal 3D nnU-Net model achieved superior performance with mean Dice scores of 35.36% (CR) and 50.22% (NR).
• This deep learning approach significantly outperformed conventional threshold-based methods (Dice scores of 15.73% for CR and 39.71% for NR).
• The models provided more accurate outcome estimates compared to existing clinical methods.

Conclusions:

• The developed deep learning models accurately estimate potential treatment outcomes for both successful and unsuccessful reperfusion in AIS.
• This advancement enables better evaluation of treatment eligibility and potential benefits, facilitating personalized treatment recommendations.
• The approach enhances clinical decision-making in AIS management by offering more precise tissue outcome predictions than traditional methods.