Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Master of Science

Program

Neuroscience

Collaborative Specialization

Machine Learning in Health and Biomedical Sciences

Supervisor

Lau, Jonathan

2nd Supervisor

Khan, Ali

Co-Supervisor

Abstract

For patients with intractable epilepsy the surgical placement of intracranial electrodes can better localize the seizure onset zone. Stereoelectroencephalography (SEEG) is one technique, where depth electrodes made of multiple contacts record activity in the brain. The precise interpretation of recordings requires the anatomical localization of each contact. Contact positions can be manually localized or determined using semi-automated algorithms. This thesis works towards the automation of SEEG contact localization with a 3D U-Net, a deep learning architecture optimized for biomedical image segmentation. The first chapter will introduce the clinical workflow for SEEG, available tools, and the potential role of deep learning. The second chapter will cover the proposed algorithm and validation methods. The last two chapters will present the accuracy of the U-Net in contrast to approaches currently employed in the clinic. Overall, the error and accuracy of the proposed method compares favorably in an independent set of clinical data. Future work will look to continue to optimize performance.

Summary for Lay Audience

Epilepsy is a neurological disorder characterized by abnormal electrical activity known as seizures. A subset of patients diagnosed with epilepsy do not respond to medication. For some, surgical removal of a part of the brain determined to be the “seizure onset zone” can significantly reduce the severity and number of seizures and can be curative. Potential candidates for surgery are determined during a pre-surgical evaluation, where diagnostic tests aim to determine any structural deficits or abnormal brain activity. When non-invasive methods fail to localize the seizure onset zone, the implantation of electrodes directly within the brain can be used to record electrical activity. Stereoelectroencephalography (SEEG) is one method using depth electrodes made of multiple contacts to simultaneously study multiple deep brain regions. Analysis of recordings requires the anatomical position of each contact in the brain, as viewed in post-operative imaging. This localization process is typically done manually or with semi-automated tools, in a process that can be both time consuming and labor intensive. The thesis works towards the automated localization of SEEG contacts using a U-Net, a deep learning model that is used in most computer vision tasks with medical images. The first chapter covers the clinical workflow, the prerequisite knowledge for medical imaging, image processing, and deep learning, and the current tools for SEEG contact localization. The second chapter will outline the implementation of the U-Net model and how the model will be validated. Finally, it will compare errors to the current tool employed in the clinic. Overall, the error and accuracy of the proposed method compares favorably in an independent set of clinical data, while reducing the amount of manual intervention required. Future work will look to continue to optimize performance.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS