Acute ischemic strokes are a major cause for death and severe neurologic deficits in the western hemisphere. The prediction of tissue outcome in case of an acute ischemic stroke is an important variable for treatment decision. An estimation of the expected outcome is typically obtained by thresholding a single perfusion parameter map, which is calculated from a perfusion CT dataset. However, cerebral perfusion is complex and the severity of perfusion impairment is not consistent within the penumbra of an acute ischemic stroke. Therefore, the application of only one parameter for acute stroke tissue outcome prediction may oversimplify the given problem. The aim of this study was to develop and evaluate the feasibility of a multiparametric approach for estimating tissue outcome in acute ischemic stroke patients using 15 CT perfusion datasets. For this purpose, perfusion parameter maps of cerebral blood flow, cerebral blood volume and mean transit time were calculated based on the concentration time curves derived from perfusion CT datasets. The parameter maps of ten patients were employed for a voxel-wise training of a support vector machine using ground-truth final infarct segmentations, whereas the remaining five patient datasets were used for evaluation of the voxel-wise prediction of tissue outcome using the trained support vector machine. Furthermore, tissue outcome was also predicted by optimal thresholding of corresponding time-to-peak (TTP) maps for comparison purposes. Both predictions were compared to ground-truth final infarct lesions for the five datasets used for evaluation. The proposed multiparametric tissue outcome prediction lead to superior prediction results in all cases. More precisely, the multiparametric prediction lead to a mean Dice coefficient of 0.556, while optimal thresholding of TTP maps lead to an average Dice-coefficient of 0.444 compared to the ground-truth infarct lesions. In conclusion, the evaluation results of the proposed method suggest that a multiparametric tissue outcome prediction may be feasible for CT perfusion datasets but needs to be evaluated in more detail.
|