Compartment Model-Guided Unsupervised Reconstruction for Low-Count Dynamic PET Imaging

Image reconstruction from low-count dynamic Positron Emission Tomography (PET) data is challenging due to the ill-conditioning of the inverse problem and poor statistics. This paper introduces a novel compartment model-guided unsupervised method that captures spatio-temporal correlations between images. By predicting tracer concentrations in each compartment, we generate prior images projected into the sinogram domain for comparison with measured data. We integrate physiological, boundary, and data constraints into the loss function while updating kinetic and network parameters to ensure accurate solution propagation. Our method enhances image quality under low-count conditions and in early frames, effectively combining physiological and physical constraints. This approach improves interpretability and generalizability while eliminating the need for training datasets. Experimental results demonstrate its superior performance compared to other reconstruction techniques.