Toward robust modeling of breast biomechanical compression: an extended study using graph neural networks

Area of Science:

• Medical imaging
• Computational modeling
• Biomedical engineering

Background:

• Accurate simulation of breast tissue deformation is crucial for image registration between 3D modalities and 2D mammograms.
• Finite element analysis (FEA) offers high-fidelity but is computationally intensive, limiting its use in rapid simulations.
• Physics-based graph neural networks (PhysGNN) present a computationally efficient alternative for modeling breast compression.

Purpose of the Study:

• Evaluate PhysGNN performance on new digital breast phantoms.
• Assess the impact of training PhysGNN on multiple phantoms.
• Extend prior work on PhysGNN for breast deformation simulation.

Main Methods:

• PhysGNN was trained on single-phantom (per-geometry) and multiphantom (multigeometry) datasets.
• Datasets were generated from incremental FEA simulations of digital breast phantoms.
• A leave-one-deformation-out strategy evaluated predictive performance under compression.

Main Results:

• PhysGNN demonstrated robust performance on new digital phantoms, with some variability due to anatomical diversity.
• Multiphantom training improved robustness and reduced prediction errors compared to single-phantom training.
• Model performance remained strong with per-geometry training, enhanced by multigeometry approaches.

Conclusions:

• PhysGNN provides a computationally efficient alternative to FEA for simulating breast compression.
• Multigeometry training enhances PhysGNN's predictive accuracy and robustness.
• PhysGNN shows significant potential for developing reliable models for compressed breast volumes, aiding image registration and algorithm development.