Dynamic PET reconstruction using the kernel method with non-local means denoising

Abstract

Non-local means with a spatiotemporal search window (NLM-ST) has been developed to denoise dynamic positron emission tomography (PET) images. The improved image quality, however, may not be good enough to generate reliable parametric images. In this work, we propose an iterative reconstruction algorithm which aims to improve the quality of dynamic PET images by incorporating NLM-ST denoising directly within the kernelized expectation-maximization (KEM) reconstruction algorithm. Since the NLM-ST denoising was employed after each KEM update, the proposed algorithm was called NLM-ST-AU-KEM. Computer simulations were conducted to evaluate the performance of the proposed reconstruction algorithm, and the results were compared to KEM with a post-reconstruction NLM-ST denoising filter (KEM + NLM-ST). The root mean squared errors (RMSE) of the dynamic PET images reconstructed using the KEM algorithm were increased after 40 iterations. Both the NLM-ST-AU-KEM and the KEM + NLM-ST methods could achieve stable RMSE values after 50 iterations, but the former had lower RMSE values. Compared to the proposed NLM-ST-AU-KEM method, the KEM + NLM-ST method tended to over-smooth dynamic PET images and parametric images. For K1 and Ki, the proposed NLM-ST-AU-KEM method had lower bias but higher variance than the KEM + NLM-ST method. For k2 and k3, the proposed NLM-ST-AU-KEM method had higher variance than the KEM + NLM-ST method, but the higher variance could be reduced by applying a kernel-based post-filtering method to the NLM-ST-AU-KEM-generated parametric images. NLM-ST denoising during image reconstruction seems to be a better strategy than that after image reconstruction.