Impact of the reperfusion status for predicting the final stroke infarct using deep learning

International audiencePredictive maps of the final infarct may help therapeutic decisions in acute ischemic stroke patients. Our objectives were to assess whether integrating the reperfusion status into deep learning models would improve their performance, and to compare them to current clinical prediction methods. Methods: We trained and tested convolutional neural networks (CNNs) to predict the final infarct in acute ischemic stroke patients treated by thrombectomy in our center. When training the CNNs, non-reperfused pa-tients from a non-thrombectomized cohort were added to the training set to increase the size of this group. Baseline diffusion and perfusion-weighted magnetic resonance imaging (MRI) were used as inputs, and the lesion segmented on day-6 MRI served as the ground truth for the final infarct. The cohort was dichotomized into two subsets, reperfused and non-reperfused patients, from which reperfusion status specific CNNs were developed and compared to one another, and to the clinically-used perfusion-diffusion mismatch model. Evaluation metrics included the Dice similarity coefficient (DSC), precision, recall, volumetric similarity, Hausdorff distance and area-under-the-curve (AUC). Results: We analyzed 109 patients, including 35 without reperfusion. The highest DSC were achieved in both reperfused and non-reperfused patients (DSC =0.44 ±0.25 and 0.47 ±0.17, respectively) when using the corresponding reperfusion status-specific CNN. CNN-based models achieved higher DSC and AUC values compared to those of perfusion-diffusion mismatch models (reperfused patients: AUC =0.87 ±0.13 vs 0.79 ±0.17, P <0.001; non-reperfused patients: AUC =0.81 ±0.13 vs 0.73 ±0.14, P