Informatics in Medicine Unlocked (Jan 2022)
Validation of FEA-based breast deformation simulation using an artificial neural network
Abstract
Background and objectives:: Imaging of breast tumors is a difficult problem because tumors shift when patients change their positions. In addition, the images taken at the diagnostic stage do not correspond to the actual breast shape during treatment. Currently, finite element analysis (FEA) is the most accurate way to simulate breast deformation. As FEA is computationally expensive, we propose to use an Neural Network (NN) to learn the deformation and use the NN during treatment. In our previous study [1], we have shown that an NN is accurate in predicting breast nodal displacements when comparing the NN result with FEA simulation. In this research, we used a breast phantom to validate NN result. Methods:: In this paper, we did magnetic resonance imaging (MRI) of a breast phantom in several positions, implemented an iterative approach to find the material properties, and simulated multiple breast deformations at different positions in FEA. The nodal displacements were used to train an NN that takes gravity direction as the input and produces the nodal displacements of surface points and the tumor center position. Validation was performed with FEA, NN, and the breast phantom MRI. Results:: Our FEA simulation with best material estimation results in an average closest point-to-point (CPTP) error of 3.63 ± 1.12 mm when the angle is 25°. The NN’s average point-to-point (PTP) error is 0.19 ± 0.12 mm when compared to FEA simulation in the validation set. The CPTP error between the NN and the MRI reconstructed surface points at 50° is 2.69 ± 0.7 mm. Our work shows that NN is a promising technique to help doctors visualize breast deformation and tumor location during treatment. Conclusions:: We used a breast phantom and showed that NN is a good approach for breast deformation simulation subject to gravity. This is the first significant step towards simulating real-time deformation using preoperative images. The same approach can be extended to improve in-situ visualization by adding other external forces that come into play during breast surgery.
