Physics-Informed Implicit Neural Representations in Biomedical Tomography: Principles, Applications, and Reliability

Biomedical tomography reconstructs spatially and temporally varying images or biophysical fields from indirect measurements governed by acquisition physics. Physics-informed implicit neural representations (INRs) parameterize these unknowns as coordinate-conditioned neural fields, optionally incorporating physical states, and estimate their parameters through measurement consistency. Despite growing adoption, the specific contributions of INRs are often conflated with those of forward modeling, continuous representation, and neural priors, obscuring when they offer substantive advantages. This review adopts a representation-physics-inference perspective to clarify the role of INRs relative to conventional voxel, basis, and other neural parameterizations. We examine how coordinate and state conditioning provide a direct interface to the forward model, while shared network parameters couple reconstruction updates across queried locations and states. Starting from scan-specific INR reconstruction, we identify three limitations: insufficient measurement information, modeling uncertainty, and the computational burden. Corresponding advances are organized into prior-enhanced reconstruction, joint estimation of target fields and physical variables, and scalable or amortized inference. Using representative examples from CT, MRI, ultrasound, photoacoustic imaging, PET, and optical tomography, we assess how these strategies affect reconstruction bias, identifiability, quantitative reliability, and computational efficiency. We distinguish essential, enabling, or merely incidental INR roles according to their coupling with acquisition physics. This perspective clarifies the opportunities and limitations of physics-informed INRs and identifies forward-model fidelity, solution identifiability, scalable inference, and independent quantitative and clinical validation as development priorities.