Skip to main content

Machine learning and interactive real-time simulation for training on relevant total hip replacement skills.

Aguilera Canon, M. C., 2021. Machine learning and interactive real-time simulation for training on relevant total hip replacement skills. Doctoral Thesis (Doctoral). Bournemouth University.

Full text available as:

[img]
Preview
PDF
AGUILERA CANON, Mara Catalina_Ph.D._2021.pdf
Available under License Creative Commons Attribution Non-commercial.

38MB

Abstract

Virtual Reality simulators have proven to be an excellent tool in the medical sector to help trainees mastering surgical abilities by providing them with unlimited training opportunities. Total Hip Replacement (THR) is a procedure that can benefit significantly from VR/AR training, given its non-reversible nature. From all the different steps required while performing a THR, doctors agree that a correct fitting of the acetabular component of the implant has the highest relevance to ensure successful outcomes. Acetabular reaming is the step during which the acetabulum is resurfaced and prepared to receive the acetabular implant. The success of this step is directly related to the success of fitting the acetabular component. Therefore, this thesis will focus on developing digital tools that can be used to assist the training of acetabular reaming. Devices such as navigation systems and robotic arms have proven to improve the final accuracy of the procedure. However, surgeons must learn to adapt their instrument movements to be recognised by infrared cameras. When surgeons are initially introduced to these systems, surgical times can be extended up to 20 minutes, maximising surgical risks. Training opportunities are sparse, given the high investment required to purchase these devices. As a cheaper alternative, we developed an Augmented Reality (AR) alternative for training on the calibration of imageless navigation systems (INS). At the time, there were no alternative simulators that using head-mounted displays to train users into the steps to calibrate such systems. Our simulator replicates the presence of an infrared camera and its interaction with the reflecting markers located on the surgical tools. A group of 6 hip surgeons were invited to test the simulator. All of them expressed their satisfaction with the ease of use and attractiveness of the simulator as well as the similarity of interaction with the real procedure. The study confirmed that our simulator represents a cheaper and faster option to train multiple surgeons simultaneously in the use of Imageless Navigation Systems (INS) than learning exclusively on the surgical theatre. Current reviews on simulators for orthopaedic surgical procedures lack objective metrics of assessment given a standard set of design requirements. Instead, most of them rely exclusively on the level of interaction and functionality provided. We propose a comparative assessment rubric based on three different evaluation criteria. Namely immersion, interaction fidelity, and applied learning theories. After our assessment, we found that none of the simulators available for THR provides an accurate interactive representation of resurfacing procedures such as acetabular reaming based on force inputs exerted by the user. This feature is indispensable for an orthopaedics simulator, given that hand-eye coordination skills are essential skills to be trained before performing non-reversible bone removal on real patients. Based on the findings of our comparative assessment, we decided to develop a model to simulate the physically-based deformation expected during traditional acetabular reaming, given the user’s interaction with a volumetric mesh. Current interactive deformation methods on high-resolution meshes are based on geometrical collision detection and do not consider the contribution of the materials’ physical properties. By ignoring the effect of the material mechanics and the force exerted by the user, they become inadequate for training on hand- eye coordination skills transferable to the surgical theatre. Volumetric meshes are preferred in surgical simulation to geometric ones, given that they are able to represent the internal evolution of deformable solids resulting from cutting and shearing operations. Existing numerical methods for representing linear and corotational FEM cuts can only maintain interactive framerates at a low resolution of the mesh. Therefore, we decided to train a machine-learning model to learn the continuum mechanic laws relevant to acetabular reaming and predict deformations at interactive framerates. To the best of our knowledge, no research has been done previously on training a machine learning model on non-elastic FEM data to achieve results at interactive framerates. As training data, we used the results from XFEM simulations precomputed over 5000 frames for plastic deformations on tetrahedral meshes with 20406 elements each. We selected XFEM simulation as the physically-based deformation ground-truth given its accuracy and fast convergence to represent cuts, discontinuities and large strain rates. Our machine learning-based interactive model was trained following the Graph Neural Networks (GNN) blocks. GNNs were selected to learn on tetrahedral meshes as other supervised-learning architectures like the Multilayer perceptron (MLP), and Convolutional neural networks (CNN) are unable to learn the relationships between entities with an arbitrary number of neighbours. The learned simulator identifies the elements to be removed on each frame and describes the accumulated stress evolution in the whole machined piece. Using data generated from the results of XFEM allowed us to embed the effects of non-linearities in our interactive simulations without extra processing time. The trained model executed the prediction task using our tetrahedral mesh and unseen reamer orientations faster per frame than the time required to generate the training FEM dataset. Given an unseen orientation of the reamer, the trained GN model updates the value of accumulated stress on each of the 20406 tetrahedral elements that constitute our mesh during the prediction task. Once this value is updated, the tetrahedrons to be removed from the mesh are identified using a threshold condition. After using each single-frame output as input for the following prediction repeatedly for up to 60 iterations, our model can maintain an accuracy of up to 90.8% in identifying the status of each element given their value of accumulated stress. Finally, we demonstrate how the developed estimator can be easily connected to any game engine and included in developing a fully functional hip arthroplasty simulator.

Item Type:Thesis (Doctoral)
Additional Information:If you feel that this work infringes your copyright please contact the BURO Manager.
Uncontrolled Keywords:graph networks; physically-based deformation; machine learning; acetabular reaming; material removal
Group:Faculty of Media & Communication
ID Code:35672
Deposited By: Unnamed user with email symplectic@symplectic
Deposited On:21 Jun 2021 12:34
Last Modified:21 Jun 2021 12:34

Downloads

Downloads per month over past year

More statistics for this item...
Repository Staff Only -