Computed Tomography Perfusion-Based Prediction of Core Infarct and Tissue at Risk: Can Artificial Intelligence Help...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Imaging

Background:

• Accurate assessment of core infarct and penumbral tissue is crucial for stroke management.
• Computed tomographic (CT) perfusion imaging provides essential data for this assessment.
• Current methods may be limited by vendor specificity and data completeness.

Purpose of the Study:

• To explore the feasibility of automated, vendor-neutral prediction of core infarct and penumbral tissue using neural networks.
• To evaluate the performance of models using complete and partial computed tomographic perfusion data.
• To assess the arterial input function independent prediction capabilities.

Main Methods:

• Developed and validated 2D U-net models for cerebral blood flow (CBF) and time to maximum (Tmax) maps using retrospective CT perfusion data from 57 patients.
• Trained models on full 28 time points and tested with reduced data sets (14, 10, 7 time points).
• Compared model-derived maps with ground truth using structural similarity index measure (SSIM) and Pearson correlation for volumes.

Main Results:

• CBF and Tmax maps derived using 28 and 14 time points showed similar SSIM (0.80-0.81) compared to ground truth.
• Pearson correlation coefficients for volumes were 0.69 for CBF and 0.74 for Tmax when comparing model-derived with RAPID software.
• Predicted maps demonstrated concordance in laterality with commercial perfusion maps; mean Dice scores were 0.54 (core infarct) and 0.63 (hypoperfusion).

Conclusions:

• AI model-derived volumes correlate well with commercial software (RAPID) for CBF and Tmax.
• Perfusion map quality remains similar using 14 time points versus 28, suggesting feasibility with partial data.
• Vendor-neutral AI models for CT perfusion processing using complete or partial data are feasible, with potential for optimization using larger datasets.