Limitations of simple machine-learning surrogates for cross-tissue prediction of radial177Ludose point kernels.

Abstract

Objective.To determine whether simple regularized linear regression models, relying solely on radial distance, mass density and elemental tissue composition as input features, can outperform the conventional water-kernel interpolation baseline for predicting radial dose point kernels (DPKs) ofacross a diverse set of homogeneous tissues and one additional low-density homogeneous medium.Approach.Radial DPKs forwere generated using OpenGATE 10 Monte Carlo simulations in 12 homogeneous media covering soft tissues, bone-like structures, and low-density lung-like tissues, together with one additional low-density homogeneous medium (g cm) evaluated separately as an out-of-distribution extrapolation test. Three scikit-learn models (RidgeCV, LassoCV, and ElasticNetCV) were trained with robust scaling and evaluated through leave-one-tissue-out (LOTO) and leave-one-family-out cross-validation, as well as on the independent low-density case. Performance was assessed using mean absolute percentage error on physical dose (), logarithmic root mean square error (), relative integral energy error, and relative errors onand, systematically compared against a water-kernel baseline obtained by linear resampling of a single simulated water DPK.Main results.ElasticNetCV was the best-performing linear surrogate overall, achieving a meanof 224% in LOTO validation (median 117%), followed by LassoCV and RidgeCV. However, none of the linear surrogates outperformed the water-kernel baseline on the main physical performance metrics. Mean gains relative to the baseline remained negative across tissue families, ranging from aboutpercentage points in bone-like tissues to aboutpercentage points in low-density lung-like tissues. Relative integral energy errors reached several thousand percent in lung-like media, and the separate low-density homogeneous medium also showed severe degradation (best surrogateversusfor the water-kernel baseline). Although some surrogate predictions showed lower log-domain errors than the baseline in selected cases, these improvements did not translate into better dosimetric accuracy in physical dose units.Significance.Simple linear machine-learning surrogates based on radial distance, mass density and macroscopic elemental composition did not improve, and often substantially worsened, the accuracy of radialDPK prediction compared with standard water-kernel interpolation. This negative benchmark delineates the limitations of basic tabular linear regression for capturing density- and composition-dependent kernel changes across a range of homogeneous media relevant to nuclear medicine dosimetry, and provides a reproducible reference framework for future development of more advanced surrogate models.