Kernelised EM image reconstruction for dual-dataset PET studies

In a number of applications, positron emission tomography (PET) requires two or more scans to observe and even quantify changes in function (e.g. tissue metabolism, or receptor binding potentials). Conventionally the raw datasets are reconstructed into images independently, allowing no sharing of information. Kernelised EM (KEM) is a recently proposed PET reconstruction method that utilises one or more sources of prior information to construct spatial basis functions with machine learning techniques for PET reconstruction. Feature vectors are defined for each pixel in the image, and the level of correlation between these feature vectors for different pixels allows spatial basis functions to be created that capture these correlations between similar pixels. This work proposes the use of KEM in dual-dataset PET studies, where the reconstructed image from the primary scan is used to construct spatial basis functions for subsequent scan reconstructions. Feature vectors for each pixel were defined as square patches centred on the pixel in question in the primary scan reconstructions. Using these patch-based feature vectors, basis functions were formed and used to reconstruct the subsequent datasets. Computer simulation studies of a simple dual-dataset brain tumour response study with a 50% change in uptake showed that KEM produces lower errors in the tumour and other regions of structure for the second scan when compared to conventional methods that use no prior information. In addition, visual quality of KEM reconstructions was observed to be superior to that of the competing methods. These results suggest that KEM is applicable to dual-dataset studies, improving reconstructed follow-up images compared to the conventional approaches.