Open Collections

Machine learnt treatment : machine learning and registration techniques for digitally planned jaw reconstructive surgery

University of British Columbia

The continuous advent of novel imaging technologies in the past two decades has created new avenues for biomechanical modeling, biomedical image analysis, and machine learning. While there still is relatively a long way ahead of the biomedical tools for them to be integrated into the conventional clinical practice, biomechanical modeling and machine learning have shown noticeable potential to change the future of treatment planning. In this work, we focus on some of the challenges in the modeling of the masticatory (chewing) system for the treatment planning of jaw reconstructive surgeries. Here, we discuss novel methods to capture the kinematics of the human jaw, fuse information in between imaging modalities, estimate the missing parts of the 3D structures (bones), and solve the inverse dynamics problem to estimate the muscular forces. This research is centered around the human masticatory system and its core component, the mandible (jaw), while focusing on the treatment planning for cancer patients. We investigate jaw tracking and develop an optical tracking system using subject-specific dental attachments and infrared markers. To achieve that, a fiducial localization method was developed to increase the accuracy of tracking. In data fusion, we propose a method to register the 3D dental meshes on the MRI of the maxillofacial structures. We use fatty ellipsoidal objects, which resonate in MRI, as fiducial landmarks to automate the entire workflow of data fusion. In shape completion, we investigate the feasibility of generating a 3D anatomy from a given dense representation using deep neural architectures. We then extend on our deep method to train a probabilistic shape completion model, which takes a variational approach to fill in the missing pieces of a given anatomy. Lastly, we tackle the challenge of inverse dynamics and motor control for biomechanical systems where we investigate the applicability of reinforcement learning (RL) for muscular force estimation. With the mentioned portfolio of methods, we try to make biomechanical modeling more accessible for clinicians, either via automating known manual processes or introducing new perspectives.

Text

2020-06-25

Attribution-NonCommercial-ShareAlike 4.0 International

http://hdl.handle.net/2429/74787

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-sa/4.0/