Deep Learning‐Based Prediction of Final Infarct Core from CT Perfusion Data: A Comparison to the Clinical Standard

Abstract

Background In the management of acute ischemic stroke, computed tomography perfusion (CTP) is used to define the ischemic core and penumbra to estimate tissue fate after reperfusion therapy. The core/penumbra dichotomy uses single‐value thresholds, which potentially discards valuable data and oversimplifies the complexity of core and penumbral estimation. Advancing the dichotomous CTP output to a probability model has several advantages such as more sophisticated modeling of pathophysiology, supporting reader interpretation, and assessing a greater range of available data to estimate tissue fate. Methods In this retrospective study, we developed a CTP probability model to move away from single perfusion thresholds to estimate tissue fate. All patients from the International Stroke Perfusion Imaging Registry database had baseline CTP and were included in the current study if they had a large vessel occlusion that recanalized fully after thrombectomy and had follow‐up diffusion‐weighted imaging. Data were split into training, validation, and testing groups. Training and validation cohorts were used to develop a deep learning model in project MONAI (Medical Open Network for Artificial Intelligence) and performance metrics were derived from the testing set. Results In total, 243 patients were included in the study. The Attention U‐Net was the best performing deep learning model, producing the best prediction of follow‐up infarct core on the test set (n = 48): mean diverse counterfactual explanations score = 0.430±0.213, mean area under the curve = 0.765±0.095; better than the single‐value thresholding with a diverse counterfactual explanations scoreof 0.247±0.167 (paired t ‐test, P <0.0001) and area under the curveof 0.604±0.074 ( P <0.0001). Conclusion The deep learning probabilistic CTP model outperformed the current clinical standard, providing a more accurate core estimate than single‐threshold‐based measures.