Learning-based estimation of the unloaded state for biomechanical models of the breast

Breast deformation simulations typically require estimating a stress-free (unloaded) state, often using iterative
FEM-based methods on gravity-loaded MRI data—an approach that is computationally intensive. This work
aims to accelerate that process using a machine learning (ML) model. Building on our previous work in this
field, we adapt and extended the approach to the given problem and furthermore treat tissue stiffness as a free
parameter. After normalizing the data and extracting features from the biomechanical model, we used XGBoost
to predict the unloaded node positions from those of the gravity-loaded state. The model was trained on 418
datasets from 23 patients, each simulated with 10 different Young’s moduli. In 10-fold cross-validation, the
average prediction error was 2.7, mm compared to FEM ground truth. This ML approach reduces unloaded
state estimation time from 20 minutes to just 0.06 seconds, enabling real-time deformation.