Detection Of Arterial Input Function From Cerebral Perfusion Using DSC-MRI Based On Clustering Analysis

Dynamic susceptibility contrast-magnetic resonance imaging using endogenous contrast agent has been widely applied to cerebral perfusion weighted imaging (PWI), which can be used to quantify important physiologic parameters related to cerebral hemodynamic, including cerebral blood flow (CBF), cerebral blood volume (CBV) and mean transit time (MTT). DSC-MRI has been playing more and more important role in clinical disease diagnosis and choice of suitable therapy. Hence, rapid, accurate and robust quantification of DSC-MRI is very significant for clinical practice.When quantifying the cerebral hemodynamic based on the technique of DSC-MRI, the knowledge of a reliable arterial input function (AIF) is required in advance. The accuracy of AIF will have a substantial influence on final calculation of cerebral hemodynamic. Traditionally, the AIF was selected by the experienced radiologist through picking several arterial voxels manually in the scanned slice covering middle cerebral artery (MCA) or internal carotid artery (ICA). However, this kind of manual method is time-consuming and dependent on the experience of operators, which resulted in the irreproducibility between different operators or the same operator at different time points. Meanwhile, because of relatively low spatial resolution for DSC-MRI image, the result of AIF derived from manual method will be contaminated by partial volume effect (PVE) severely. Hence, it is still a urgent and practical problem to reduce human intervention during AIF detection by developing an automatic or semi-automatic algorithms.In order to solve the disadvantage of manual method for AIF detection, we evaluated the performances of different clustering algorithms for cerebral AIF detection. This work included the following steps:acquire the PWI images of forty-two healthy volunteers using the technique of DSC-MRI; correct for motion and rotation of the volume images at different time points caused by breathe, heartbeat and some other involuntary factors based on an off-line workstation; select the scanned slice in the first volume image covering the MCA by browsing all the slices manually; convert the time-intensity curves of image signal in the previously selected slice into time-concentration curves of contrast agent; exclude the curves with the smallest areas, severe fluctuation and several PVE contamination; finally, several different kinds of clustering methods were applied to the residual curves to detect the AIF voxels automatically. In order to obtain the best method, we compared the detection accuracy, calculation-recalculation reproducibility and computational complexity during the automatic detection of AIF. Because of lack of support for golden standard in clinical study, we also added the simulation study to our work which supplied the true AIF. The feasibility of every clustering analysis for AIF detection was evaluated by comparing the estimated AIF and the true AIF.Our results demonstrated that, (1) for irreproducible clustering algorithms, compared with traditional manual method, k-means method can obtain more accurate AIF; relative to fuzzy c-means method, k-means clustering analysis can obtain AIF with more accuracy and better robustness; (2) for reproducible clustering algorithms, relative to fast affine propagation (FastAP), normalized cut (Ncut) clustering can obtain more accurate AIF as well as agglomerative hierarchical (AH) clustering. In addition, the computational complexity of Ncut algorithm during AIF detection was lower than that of AH method. Hence, we can come to the conclusion that Ncut algorithm indicated a better future prospect in AIF detection.