Predicting trabecular arrangement in the proximal femur: An artificial neural network approach for varied geometries and load cases

Machine learning (ML) and deep learning (DL) approaches can solve the same problems as the finite element method (FEM) with a high degree of accuracy in a fraction of the required time, by learning from previously presented data. In this work, the bone remodelling phenomenon was modelled using feed-forward neural networks (NN), by gathering a dataset of density distribution comprising several geometries and load cases. The model should output for some point in the domain the its apparent density, taking into consideration the geometric and loading parameters of the model . After training. the trabecular distribution obtained with the NN was similar to the FEM while the analysis was performed in a fraction of the time, having shown a reduction from 1020 s to 0.064 s. It is expected that the results can be extended to different structures if a different dataset is acquired. In summary, the ML approach allows for significant savings in computational time while presenting accurate results.