Thesis Format

Integrated Article

Automated Segmentation of Temporal Bone Structures

Daniel Allen, The University of Western OntarioFollow

Degree

Master of Engineering Science

Program

Electrical and Computer Engineering

Supervisor

Ladak, Hanif M.

2nd Supervisor

Agrawal, Sumit K.

Co-Supervisor

Abstract

Mastoidectomy is a challenging surgical procedure that is difficult to perform and practice. As supplementation to current training techniques, surgical simulators have been developed with the ability to visualize and operate on temporal bone anatomy. Medical image segmentation is done to create three-dimensional models of anatomical structures for simulation. Manual segmentation is an accurate but time-consuming process that requires an expert to label each structure on images. An automatic method for segmentation would allow for more practical model creation. The objective of this work was to create an automated segmentation algorithm for structures of the temporal bone relevant to mastoidectomy. The first method explored was multi-atlas based segmentation of the sigmoid sinus which produced accurate and consistent results. In order to segment other structures and improve robustness and accuracy, two convolutional neural networks were compared. The convolutional neural network implementation produced results that were more accurate than previously published work.

Summary for Lay Audience

Surgeries in the area of the ear can be difficult to train and practice. There are many small important structures to be considered and there is a lot of variation between patients. If mistakes are made during the surgery, it can cause severe damage to the patient. By using a surgical simulator, surgical trainees can improve their skills before operating on real patients at a much lower cost than when using classical training methods. To create a surgical simulator, anatomical structures need to be labeled from images so that 3D models can be made. This is called image segmentation and can be done manually or automatically. Manual labelling is very accurate but takes a long time and requires an expert to do it. Automatic labelling is much easier and faster to do in a clinical setting. However, many parts of the anatomy that need to be labeled are small, variable in position and shape, and have low contrast edges (hard to distinguish from surrounding objects). These issues make automating the labelling of the structures very difficult. This work compares multiple methods for automatic image labelling. The first method tested and developed was done on the sigmoid sinus, a vein that passes near the ear. A set of high-resolution manually labelled examples of the vein were used and transformed to match the new lower resolution images to be labelled, and then these sets were combined. This method resulted in labels that were similar to the actual labels. The second method was done on several anatomic structures of the ear and used deep learning networks to learn patterns in the images and label them automatically. This method quickly and successfully created automatic labels from images that were also very close to the actual labels and showed better results than previous work on the same structures. This labelling method may be used to create 3D models for surgical simulators.

Recommended Citation

Allen, Daniel, "Automated Segmentation of Temporal Bone Structures" (2019). Electronic Thesis and Dissertation Repository. 6605.
https://ir.lib.uwo.ca/etd/6605