Author: Daniela Mini
Mini, Daniela, 2024 Computationally efficient assessment of bone strain for a proximal humeral fracture with a fracture fixation plate, Flinders University, College of Science and Engineering
Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.
The treatment of proximal humeral fractures with fracture fixation plates has been associated with a failure rate of up to 35% and the cause of this is not yet fully understood. However, several factors contribute to the failures, such as fracture configuration, bone quality, screw orientation, and screw length. Finite Element (FE) analysis is a commonly used computational approach to investigate the biomechanics of fracture fixation devices. Still, FE techniques are too computationally expensive for complex problems with a high number of variables. The overall aim of this thesis is to develop and assess computationally efficient methodologies in order to investigate the bone deformation of a fractured humerus with a fracture fixation plate, varying different implant parameters and within subjects. The first study of this thesis aimed to develop a computational model to predict the minimum principal strain within the fractured humerus by varying the length of the proximal screws. To achieve this, an FE analysis and an Adaptive Neural Network (ANN) method were combined. A semi-automated FE workflow on a single subject was developed to generate up to a thousand different configurations by varying the length of the proximal screws in the humeral head. The data generated were used to train ANN models, which showed a high level of accuracy (R2 = 0.96-0.99, RMSE = 0.51-1.27 % strain). After the training process, the best ANN model was applied to predict bone strain for a full factorial scenario. This confirmed that the length of the cortical screw has a significant impact on bone strain. Additionally, using the ANN to make predictions for the full factorial scenario only took a few seconds, whereas performing FE analysis for the same number of configurations would have required 170.6 days.
The second study aimed to develop an efficient computational model to reproduce the minimum principal strain within the fractured humerus with the variation of the orientation of the proximal screws. Similar to the first study, a semi-automated pipeline was developed to generate FE models based on a single subject, with varying orientations of the proximal screws, resulting in up to a thousand simulations. The data from the models created were used to train ANN models. In particular, two types of ANN models were developed, one to predict the screw collision and the other to predict the strain around the screws when the orientation of the screws varies. The two models combined had a good level of accuracy, showing a percentage of error of 15.60% for the first model, and an R2 = 0.98-0.99 and an RMSE = 0.65-3.77 % strain for the second model. The two ANN models were then used to make predictions of a full factorial scenario, showing again the high impact of the cortical screw. Moreover, using the ANN for predictions in a full factorial scenario took only a few seconds, compared to the FE analysis which would have taken approximately 5,911.8 days.
Despite the high efficiency of ANN models, they have the limitation of predicting a single value of strain, instead of the distribution of bone strain around the surface of the screws. For this reason, a more advanced DL technique has been introduced in the last two studies.
In the third study, the FE data generated from the first and second analyses were used to train two types of Graph Neural Network (GNN) models. The models were trained using nodal information from the bone surface of proximal screws from the FE data and could predict the distribution of bone strain around the screws, with an R2 = 0.87-0.95 and RMSE = 2.81-3.86 % strain. The GNN models showed a high level of accuracy, and the main advantage was the considerably smaller training and testing time, consisting of respectively a few hours and a few seconds.
The last study aimed to generate an efficient computational model on a cohort of subjects. A semi-automated FE workflow was developed to generate data from 434 subjects. These data were used to train a GNN model, which could predict the distribution of minimum, middle and maximum principal strain around the proximal screws. The results of the GNN model were satisfactory, with an R2 = 0.76 and an RMSE = 4.93 % strain for the prediction of the minimal principal strain of the bone, showing again the time efficiency of GNN models.
This thesis has successfully developed methodologies with different levels of complexity that combine FE analysis with DL techniques to enhance and expedite the computational process. Specifically, it has been shown to be time-efficient without compromising the accuracy of the FE analysis. Integrating DL algorithms into FE setups for evaluating medical device performance has the potential to improve surgical planning for individual patients, ultimately leading to better outcomes in medical procedures.
Keywords: Finite Element Analysis, locking plate fixation, Neural Networks, proximal humerus fracture, surrogate modelling, Deep Learning
Subject: Medicine thesis
Thesis type: Doctor of Philosophy
Completed: 2024
School: College of Science and Engineering
Supervisor: Mark Taylor