Zero-Shot Spatio–Temporal Interpolation Using Unsupervised Learning for CT Perfusion With Increased Sampling Interval

Abstract

Computed tomography perfusion (CTP) is recognized as a valuable tool for acute stroke diagnosis and management. Increasing the temporal sampling interval is a simple way to reduce radiation exposure, but it results in low temporal resolution that may fail to capture rapid perfusion change. To address this issue, we proposed a novel zero-shot spatio–temporal interpolation method that used unsupervised learning to restore CTP images with increased temporal sampling interval. Real brain CTP data with a 0.5-s sampling interval were quarterly subsampled to simulate a long sampling interval (i.e., 2-s) protocol. Then, the one-quarter subsampled CTP images were restored to the complete CTP images by using the proposed method. Finally, perfusion parameters were calculated from the complete and restored CTP images. Compared to the temporally subsampled and linearly interpolated (LI) CTP images, the unsupervised learning-restored CTP images produced perfusion maps more closely aligned with those derived from the completed CTP images, in terms of structural similarity index, mean absolute difference, and contrast-to-noise ratio. Moreover, the means relative differences obtained from the proposed method were lower than those obtained from the LI method in all cases. Our preliminary study shows that using unsupervised learning to restore CTP image frames from the CTP scan with an increased temporal sampling interval is feasible, indicating the possibility of reducing the radiation dose of CTP scans.