Finite element-machine learning surrogates for synthetic recovery of breast tissue elastic properties: Observation...

This study developed and systematically evaluated a computational framework combining finite element modeling (FEM) and machine learning for the synthetic recovery of tissue-specific elastic properties from high-dimensional displacement fields generated within a fixed finite element model. A heterogeneous three-dimensional breast model comprising skin, adipose, and glandular tissues was reconstructed from medical images and used to generate 1000 finite element simulations with varying tissue elastic moduli and prescribed compression displacements. Random Forest regression models were trained to estimate the elastic modulus of each tissue from the nodal displacement field and compression displacement. The framework was evaluated through input-configuration and prediction-strategy comparisons, low-dimensional baselines, learning curve analysis, boundary-condition-free surface observations, repeated noise and spatial-subsampling experiments, structured observation degradation, marginal extrapolation, and joint parameter-space distribution-shift experiments. The tissue-specific Random Forest models achieved mean absolute percentage errors of 2.93%, 1.86%, and 1.45% for skin, adipose tissue, and glandular tissue, respectively, under five-fold cross-validation. The apparent influence of additional non-target tissue moduli was sensitive to the treatment of scalar predictors within the high-dimensional feature space, while target-wise standardization substantially improved adipose and glandular performance in the simultaneous multi-output model. Comparison with lower-dimensional baselines showed that supervised partial least squares regression outperformed the full-field Random Forest for adipose and glandular tissue, whereas the Random Forest retained lower percentage error for skin. Marginal leave-one-range-out experiments further showed that the severe extrapolation deterioration observed with Random Forest was strongly algorithm dependent: PLS substantially reduced prediction errors in all held-out target intervals. Complementary four-dimensional block holdout experiments showed model- and tissue-dependent performance under structured joint parameter-space distribution shift, with PLS providing lower errors for adipose and glandular tissue and Random Forest showing lower aggregate error for skin. Prediction accuracy also improved progressively with increasing training set size. Removing displacement components prescribed directly by the boundary conditions produced negligible changes in performance, while useful predictive information was retained using only external skin-surface observations. Repeated random and spatially structured degradation experiments, including spatially correlated noise and distributed surface-observation patterns, showed progressive, tissue-dependent performance losses with increasing noise and decreasing observation density. Overall, the results support the feasibility of machine learning-based synthetic recovery of tissue-specific elastic properties within the controlled computational setting considered in this study, while showing that model representation, target scaling, and the geometry of distribution shift substantially affect predictive behavior. These results characterize surrogate parameter recovery within the represented simulation framework and should not be interpreted as demonstrating unrestricted global identifiability or universal extrapolation capability.