• Login
• Admin

Khalili Zadeh Mahani, Golnar (2024) Interpretability and annotation scarcity in deep medical image segmentation. PhD thesis, University of Nottingham.

PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader Available under Licence Creative Commons Attribution. Download (9MB) Request Thesis

Abstract

This thesis presents an exploration into the challenges of medical image segmentation, particularly focusing on reliability, interpretability, and strong annotation sparcity, through the use of tailored loss functions. It tackles three primary issues: 1) enhancing segmentation output quality 2) addressing the scarcity of strong annotations, and 3) evaluating novel application of the techniques on Multiple Sclerosis (MS) lesion segmentation.

First, this thesis proposes the integration of uncertainty maps into medical image segmentation to gain insights into the confidence of a model. The accuracy of these maps in conveying true uncertainty is a subject of debate. This thesis tackles the intricacies of interpreting uncertainty maps in medical image segmentation, considering factors like lesion shape, border irregularity, inconsistencies in acquisition methods, variance in lesion texture, and the presence of image artifacts. The experiments explore the interpretation of these maps and evaluate their reliability in different domains. The proposed tailored loss functions incorporate model uncertainty, error feedback, and their combinations with uncertainty inherent from the data as a regularisation term.

Second, the thesis addresses the challenge of limited accurate annotations by introducing a tailored loss function designed for weakly-supervised medical image segmentation that relies solely on bounding box annotations. This novel loss function leverages predictive uncertainty estimates during training to direct model learning towards areas of the image with high confidence. It also integrates a local spatial constraint within the loss function to ensure consistent predictions within a given area. The proposed approach outperforms both state-of-the-art learning-based and non-learning-based methods, showcasing its superior efficacy.

Third, the thesis examines the use of proposed uncertainty-aware loss functions in the fully supervised segmentation of MS brain lesions. Despite progress, the challenge of accurately segmenting MS lesions and minimising misdiagnosis remains, emphasising the necessity for improved imaging biomarkers and robust segmentation techniques to increase diagnostic accuracy and patient care in MS. Experiments evaluating two preprocessing methods for extracting brain tissue demonstrate the influence of preprocessing choices on the outcomes of deep segmentation methods in complex datasets. Furthermore, we find that loss functions incorporating model uncertainty feedback using data uncertainty as a regularisation term enhance segmentation performance on unseen MS data acquired from different sites and scanners. This highlights the importance of developing customised loss functions to improve model robustness, in MS lesion segmentation models.

Item Type:	Thesis (University of Nottingham only) (PhD)
Supervisors:	French, Andrew Chen, Xin Evangelou, Nikolaos Sotiropolous, Stamatios
Keywords:	deep medical image segmentation, Diagnostic imaging--Data processing, image processing
Subjects:	Q Science > QA Mathematics > QA 75 Electronic computers. Computer science R Medicine > RC Internal medicine W Medicine and related subjects (NLM Classification) > WN Radiology. Diagnostic imaging
Faculties/Schools:	UK Campuses > Faculty of Science > School of Computer Science
Item ID:	78836
Depositing User:	Khalili Zadeh Mahani, Golnar
Date Deposited:	13 Dec 2024 04:40
Last Modified:	13 Dec 2024 04:40
URI:	https://eprints.nottingham.ac.uk/id/eprint/78836

Actions (Archive Staff Only)

Edit View