PhyConform-Nutri: Physically Constrained Joint Conformal Prediction for Nursing-Oriented Safe Carbohydrate and Energy Estimation of Potato Dishes
Аннотация
Accurate carbohydrate and energy estimation from images of potato dishes is important for automated dietary assessment and nursing-oriented dietary monitoring, yet conventional deep regression models usually provide only point predictions and cannot determine whether an estimate is reliable enough for automatic reporting. We propose PhyConform-Nutri, a lightweight framework that integrates physically constrained multi-task regression with joint conformal prediction for uncertainty-aware nutrition estimation from a single overhead RGB image. The model jointly predicts total dish mass, potato mass, total carbohydrate, potato-attributed carbohydrate, and energy through a differentiable feasible-domain parameterization, thereby enforcing non-negativity, ingredient-containment, carbohydrate-containment, and energy-density constraints by construction. To model image-dependent and correlated uncertainty, the framework learns heteroscedastic marginal scales and an inter-target correlation structure, which are used to construct a five-dimensional Mahalanobis split-conformal prediction set. The calibrated set is intersected with the nutritional feasible domain and projected onto the carbohydrate–energy plane, while a selective prediction mechanism further uses conformal-region size and feature-space out-of-distribution distance to identify cases requiring nursing or dietitian review. Experiments on a potato-focused subset of Nutrition5k demonstrate that PhyConform-Nutri improves point-estimation accuracy, maintains reliable empirical joint coverage, and generates more efficient uncertainty regions than conventional conformal prediction methods.