Dynamic positron emission tomography (PET) with [18F]FDG enables quantitative analysis of tissue metabolism through kinetic modelling. Conventional voxel-wise estimation methods based on non-linear curve fitting are computationally intensive, while data-driven deep neural networks require large training datasets and memory resources. To counter these, we present a physiological neural representation that employs implicit neural representations (INRs) for continuous, patient-specific tracer kinetic parameter estimation in dynamic PET [1]. We predict the parameters K1, K2, K3, and Vb [2] of an irreversible two-compartment model minimising the mean-squared error between modelled and measured time-activity curves. The dataset included dynamic [18F]FDG PET/CT studies of 24 oncological patients, the INRs were trained per patient, including models augmented with voxel-wise CT Hounsfield units or feature embeddings from a 3D CT foundation model [3]. These anatomical descriptors acted as implicit regularisers, helping robustness. Across the patients, we observed statistically significant reductions in modelling errors as well as computation time. Kinetic maps of K1, K2, K3, and Vb were consistent across liver, spleen, lungs, and kidneys when compared with organ-specific ranges from Sari et al. [4]. The method provides a continuous, data-efficient approach for tracer kinetics estimation, adaptable to anatomy and multi-modal physiological modelling.