Spatially Adaptive Temporal Smoothing for Reconstruction of Dynamic PET and Gated SPECT Images

In this paper we propose a method for spatio-temporal reconstruction of dynamic or gated image sequences. In a previous method we proposed, temporal smoothing in a Karhunen-Loève (KL) transform domain was used prior to reconstruction to reduce the effect of noise. Unlike the Bayesian priors that are usually used in image reconstruction, temporal KL smoothing is a data-driven approach that takes advantage of the fact that the desired part of the observations is characterized by strong interframe correlations, whereas the noise is entirely uncorrelated. In this paper we improve on our previous technique by making the temporal smoothing adapt spatially to local image characteristics. This substantially improves the noise performance of the temporal smoothing, while significantly lessening the possibility of signal distortion. In the proposed method, spatial regions of the projection-data sequence having similar statistical characteristics are identified by an unsupervised k-means clustering algorithm. A different Karhunen-Loève (KL) transformation is designed for each image region, adapting the smoothing to the local temporal behavior. Finally, images are reconstructed from the smoothed projections by existing approaches. Experimental computer simulation results are shown that demonstrate potential improvements in image quality obtained by this technique in dynamic and gated imaging applications in brain and heart. Spatially Adaptive Temporal Smoothing for Reconstruction of Dynamic PET and Gated SPECT Images Jovan G. Brankov, Yongyi Yang, Manoj V. Narayanan, and Miles N. Wernick Dept. of Electrical & Computer Engineering, Illinois Institute of Technology, Chicago, IL, USA Dept. of Nuclear Medicine, University of Massachusetts Medical Center, Worcester, MA, USA * This research was supported by the National Institute of Neurological Disorders and Stroke (NINDS) under grant NS35273. Abstract In this paper we propose a method for spatio-temporal reconstruction of dynamic or gated image sequences. In a previous method we proposed, temporal smoothing in a Karhunen-Loève (KL) transform domain was used prior to reconstruction to reduce the effect of noise. Unlike the Bayesian priors that are usually used in image reconstruction, temporal KL smoothing is a data-driven approach that takes advantage of the fact that the desired part of the observations is characterized by strong inter-frame correlations, whereas the noise is entirely uncorrelated. In this paper we improve on our previous technique by making the temporal smoothing adapt spatially to local image characteristics. This substantially improves the noise performance of the temporal smoothing, while significantly lessening the possibility of signal distortion. In the proposed method, spatial regions of the projectiondata sequence having similar statistical characteristics are identified by an unsupervised k-means clustering algorithm. A different Karhunen-Loève (KL) transformation is designed for each image region, adapting the smoothing to the local temporal behavior. Finally, images are reconstructed from the smoothed projections by existing approaches. Experimental computer simulation results are shown that demonstrate potential improvements in image quality obtained by this technique in dynamic and gated imaging applications in brain and heart.