Background: CTP imaging may be used in cases of acute ischemic stroke (AIS) to predict response to treatment. Single-value thresholds are often applied to CTP-derived maps (e.g., Cerebral Blood Flow/CBF and Delay Time/DT) to estimate tissue fate from ischemic core and penumbra. However, limitations of using single-value thresholds for this purpose have been widely noted. Deep Learning (DL) methods have been proposed to better capture the complex pathophysiology of AIS. Here we use DL to predict tissue fate in AIS patients after successful intra-arterial thrombectomy, from CTP maps.
Methods: Patients with complete large vessel occlusions who recanalized following thrombectomy were selected from the International Stroke Perfusion Registry; acute CTP imaging and follow-up MR-DWI imaging were acquired. CTP maps were derived through MIStar (Apollo Imaging, Melbourne, Australia) and DWI final infarcts were segmented before registration to CTP. Patients were randomly split into training, validation and testing cohorts. An Attention U-Net DL network was trained to predict the follow-up infarct from DT, CBV and pre-contrast CT. Performance was evaluated over the testing set using ROC-AUC and Dice scores and compared to single-value threshold-based predictions.
Results: Altogether, 144 patients were included in the study (training/validation/testing 87/29/28). Patients underwent thrombectomy following thrombolysis (68) or thrombectomy alone (76), with Thrombolysis in Cerebral Infarction scores of 2C (62) or 3 (82). The median lesion size was 23.2 mL (min, max = [0.1, 217.3], IQR = [8.2-55.3]). Measured against DWI final infarct, the DL algorithm showed mean AUC and Dice scores of 0.69 (SD 0.12) and 0.34 (SD 0.23). MIStar ischemic core returned mean AUC and Dice scores of 0.61 (SD 0.09) and 0.24 (SD 0.15).
Discussion: The DL algorithm surpassed the ischemic core measurement in predicting tissue fate following successful thrombectomy, demonstrating the potential of DL to improve on current standards. Further developments include predicting tissue fate from raw CTP time series data.