Positron emission tomography (PET) is an advanced nuclear medicine imaging technique widely used in clinical diagnostics such as neurology and oncology. In PET image reconstruction, the widespread adoption of deep learning is attributed to its potent feature extraction capabilities. However, the challenge lies in ensuring that the employed network is interpretable and rational. Additionally, addressing the intricate issue of achieving superior results with a smaller training set remains a formidable task. In this paper, we propose a novel alternating learning dual-domain reconstruction algorithm. This method combines the likelihood function based on the PET imaging model with a learnable dual-domain regularization term as a composite objective. The objective function is minimized through alternating iterations to obtain reconstructed activity image and denoised sinogram. The iterative process enhances the convergence speed by integrating residual structures, and the assurance of result convergence is facilitated through the imposition of judgment conditions. Experimental results demonstrate that our method surpasses OSEM and DeepPET in terms of SSIM and PSNR.
|