Anomaly detection in brain imaging

Kascenas, Antanas (2023) Anomaly detection in brain imaging. EngD thesis, University of Glasgow.

Abstract

Modern healthcare systems employ a variety of medical imaging technologies, such as X-ray, MRI and CT, to improve patient outcomes, time and cost efficiency, and enable further research. Artificial intelligence and machine learning have shown promise in enhancing medical image analysis systems, leading to a proliferation of research in the field. However, many proposed approaches, such as image classification or segmentation, require large amounts of professional annotations, which are costly and time-consuming to acquire. Anomaly detection is an approach that requires less manual effort and thus can benefit from scaling to datasets of ever-increasing size.

In this thesis, we focus on anomaly localisation for pathology detection with models trained on healthy data without dense annotations. We identify two key weaknesses of current image reconstruction-based anomaly detection methods: poor image reconstruction and overdependency on pixel/voxel intensity for identification of anomalies. To address these weaknesses, we develop two novel methods: denoising autoencoder and context-tolocal feature matching, respectively.

Finally, we apply both methods to in-hospital data in collaboration with NHS Greater Glasgow and Clyde. We discuss the issues of data collection, filtering, processing, and evaluation arising in applying anomaly detection methods beyond curated datasets. We design and run a clinical evaluation contrasting our proposed methods and revealing difficulties in gauging performance of anomaly detection systems. Our findings suggest that further research is needed to fully realise the potential of anomaly detection for practical medical imaging applications. Specifically, we suggest investigating anomaly detection methods that are able to take advantage of more types of supervision (e.g. weak-labels), more context (e.g. prior scans) and make structured end-to-end predictions (e.g. bounding boxes).

• Item Type: Thesis (EngD)
• Qualification Level: Doctoral
• Additional Information: Supported by funding from Canon Medical Research Europe Limited and the Engineering and Physical Sciences Research Council (EPSRC).
• Subjects: T Technology > T Technology (General)
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Pugeault, Dr. Nicolas
• Date of Award: 2023
• Depositing User: Theses Team
• Unique ID: glathesis:2023-83832
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 10 Oct 2023 09:23
• Last Modified: 10 Oct 2023 09:25
• Thesis DOI: 10.5525/gla.thesis.83832
• URI: https://theses.gla.ac.uk/id/eprint/83832