ºìÐÓÊÓƵ

Topic

Vision Transformer-based Context-Aware System for Lingual Ultrasound in Digital Health Ecosystem

Department of Electrical and Computer Engineering

Date & location

• Tuesday, November 26, 2024
• 10:00 A.M.
• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Fayez Gebali, Department of Electrical and Computer Engineering, University of Victoria  (Co-Supervisor)

• Dr. Awos Kanan, Department of Electrical and Computer Engineering, UVic (Co-Supervisor)

• Dr. Afzal Suleman, Department of Mechanical Engineering, UVic (Outside Member) 

External Examiner

• Dr. Safwan R Wshah, Department of Computer Science, University of Vermont

Chair of Oral Examination

• Dr. Tom Ruth, Department of Physics and Astronomy, UVic

Abstract

The complex nature of modern healthcare systems and the widespread distribution of healthcare infrastructure made the interoperability within healthcare information system challenging. This could pose security risks, missing data, miscommunication, in addition to the human and technical-based errors. This dissertation focuses on utilizing an advanced AI system to overcome the challenges of clinical analysis, data confidentiality, availability, and integrity. There are three main contributions of this research. First, implement TongueTransUNet, which is a well-managed architecture that utilizes a vision transformer, UNet encoder-decoder convolutional neural network, contrastive loss and quality control process supported with human-reinforcement feedback to extract tongue fingerprint. Second, design ZTCloudGuard for access control within the telehealth cloud-based eco-system between. The architecture manage users, devices, and output attributes by deriving a score to assess the mutual relationship considering semantic and syntactic analysis. Third, utilize hybrid qualitative and quantitative evaluation metrics and conduct comparative analysis to other related research. The main applications to this research are minimizing medical errors, protecting healthcare practitioners, detecting unrelated input and undesired output. An ablation study using synthetic healthcare information attributes and word2vec model was conducted to judge the model results. The outcomes showed robustness and enhancement by focusing on high-quality input and rejecting unacceptable data. If the automatic process fails or goes below a predefined threshold, an extra reinforcement verification layer is introduced to the algorithm to add manual and human feedback.