Abstract Body: Introduction: Computed tomography perfusion (CTP) is essential in the management of acute ischemic stroke, as it determines patient eligibility for reperfusion therapy. Accurate measurement of perfusion parameters requires precise identification of the regions of interest (ROIs) for arterial input function (AIF) and venous output function (VOF). Traditional methods for setting the ROIs, based on criteria such as time to peak (TTP) and high amplitude, often suffer from inaccuracies due to individual vessel variability and signal interference. To overcome these limitations, we developed an AI-driven approach that uses machine learning to detect vascular landmarks and refine ROI selection, thereby successfully generating an improved CTP map.
Methods: We used a dataset of 499 stroke patient CT cases from 2021 to 2023, with vessel landmarks manually annotated by radiologists. Of these, 50 cases were for testing, and 449 were used for training. Our system includes two components: (1) an AI model for ROIs detection and (2) an ML model for ROIs intensity validation. The AI model, based on a Unet with ResNet as the encoder, detects eight ROIs from 4D CTP images. It is optimized using mean squared error (MSE) loss to reduce the difference between predicted and ground truth (GT) landmarks. The second component, an ML model, differentiates AIF from VOF landmarks and checks intensity normality. We used XGBoost to evaluate intensity validity and a multi-layer perceptron (MLP) to verify consistency with predicted landmarks. Both models were trained using radiologist-verified data, and data augmentation was applied to improve robustness.
Results: The AI model showed an average Euclidean distance error of 4.37 mm on the test set, indicating accurate vessel localization. The ML model achieved 0.815 accuracy, 0.92 sensitivity, and 0.725 specificity in distinguishing AIF from VOF. Our method also correlated strongly with manual radiologist segmentation, with Pearson R² scores of 0.93 for CBF < 30% and 0.913 for Tmax > 6 seconds.
Conclusion: In this study, we propose a novel AI and ML-based framework for optimizing ROI selection for AIF/VOF in CTP analysis. Our approach improves vessel localization and intensity validation while producing CTP maps closely aligned with manual expert review. This method offers a promising alternative to conventional techniques, enhancing perfusion analysis precision and supporting better clinical decision-making in acute ischemic stroke management.