Physics-Informed Machine Learning in Medical Imaging, Radiotherapy, and Biophysical Modeling: Methods, Comparative Evidence, and Clinical Readiness

Physics-informed machine learning (PIML) incorporates measurement operators, governing equations, simulated data, or mechanistic assumptions into learning. Despite encouraging biomedical demonstrations, performance against a numerical solver and physical plausibility alone do not establish trustworthy predictions or clinical benefit. This critical narrative review examines medical imaging, radiotherapy, and biophysical modeling through a unified assessment of prior fidelity, integration mechanism, fair comparators, quantitative endpoints, and evidence maturity. We distinguish physics-informed neural networks, learned surrogates, neural operators, and hybrid reconstruction pipelines; critically, we separate fundamental physical constraints from uncertain constitutive laws and phenomenological biological models. We summarize representative computational and patient-informed studies, including expanded evidence from radiation-dose modeling, and show why reported improvements in gamma agreement, segmentation, flow reconstruction, or speed cannot be pooled across distinct tasks. An original decision framework identifies when different physics-AI couplings are candidates for comparison. We introduce a proposed Physics Contribution Profile that separates predictive gain, physical validity, robustness, uncertainty calibration, and lifecycle cost; a diagnostic taxonomy of failure modes; and a reproducible, domain-specific benchmark and extraction protocol. A six-level evidence classification distinguishes numerical verification from patient outcomes. Principal limitations are model discrepancy, parameter non-identifiability, correlated simulation samples, and sparse prospective evaluation. The literature synthesis is non-exhaustive; the original database-screening records were unavailable, and the proposed benchmark has not been executed. The review proposes a testable agenda: compare matched data-only, mechanistic, and hybrid baselines; quantify conservation and uncertainty; and demonstrate transfer and clinical value independently of the simulator.