Abstract: The major challenges of positron emission tomography（PET） image reconstruction are reducing the computational cost and improving the accuracy of reconstructed image.Traditional image reconstruction is performed on a large number of acquired coincidence events with uniform pixel representation.In this paper,a structure prior guided point clouds framework is proposed to represent the imaged object,which reduces computational cost while maintains the reconstruction accuracy at the same time.The proposed framework allows extremely convenient enhancement/reduction of local precision by simply adding or removing points in the reconstruction domain.A two-step strategy is employed in our framework.In the first step,points are adaptively distributed based on the anatomical prior information acquired to establish an initial estimation of activity distributions.In the second step,more points are added to the high activity concentration regions to improve the reconstruction accuracy.At the same time,because of the reduction of dimensions of image representation and system probability matrix required in iterative statistical reconstructions,the reconstruction speed is improved effectively.In this paper,experiments on digital phantom and Monte Carlo simulations are conducted with quantitative and statistical analysis and the results on real patient scan data demonstrate the clinical potential of our proposed method.