The Library
Deep learning-based medical image segmentation with limited annotation
Tools
Chen, Jingkun (2023) Deep learning-based medical image segmentation with limited annotation. PhD thesis, University of Warwick.
PDF
WRAP_Theses_Chen_2023.pdf - Unspecified Version Embargoed item. Restricted access to Repository staff only until 11 November 2025. Contact author directly, specifying your specific needs. - Requires a PDF viewer. Download (19Mb) |
Official URL: http://webcat.warwick.ac.uk/record=b3982128~S1
Abstract
Medical image segmentation is a vital component of contemporary clinical practice, playing a crucial role in achieving precise diagnoses, personalized treatment planning, and ongoing disease monitoring. The integration of deep learning into this domain has brought about significant advancements, particularly in the automation of tumor and organ segmentation. This breakthrough alleviates the burden of labor-intensive expert annotations, improving efficiency and accuracy in medical imaging analysis. However, the effectiveness of deep learning models in this context is heavily reliant on the availability of well-annotated data. A shortage of sufficient data can result in suboptimal outcomes, potentially affecting the quality of diagnoses and patient care. This thesis addresses the challenge of enhancing medical image segmentation within the constraints of annoatated data scarcity and stringent privacy regulations within medical centres. In Chapter 3, we propose an innovative approach that harnesses a substantial repository of unlabeled medical images alongside a limited set of labeled data. This unique unpaired semi-supervised medical image segmentation method leverages shared task affinities among unpaired images to enhance model training. Chapter 4 introduces a novel strategy known as pixel-level dynamic contrastive learning. This approach adaptively adjusts pixel sampling rates based on class confidence and confusion degrees, effectively optimizing the use of sparsely labeled data and enhancing segmentation accuracy. Chapter 5 tackles the challenges associated with scribble annotated data, introducing a cross-image matching technique to standardize and efficiently incorporate these data into the network training process. Chapter 6 builds upon this foundation by proposing a novel framework that combines scribble annoatated data with unlabeled data using a student-teacher approach and consistency loss, further improving performance. Finally, Chapter 7 outlines potential future research directions in the field of medical image segmentation, underscoring the significant contributions of this thesis toward advancing the state of the art in medical image analysis and healthcare delivery. In essence, this thesis has the potential to change how medical professionals use deep learning in their daily work, benefiting patients and healthcare systems.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | R Medicine > RC Internal medicine | ||||
Library of Congress Subject Headings (LCSH): | Diagnostic imaging -- Data processing, Imaging systems in medicine, Pattern perception, Image analysis, Diagnostic imaging, Deep learning (Machine learning) | ||||
Official Date: | September 2023 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Warwick Manufacturing Group | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Zhang, Jianguo, Han, Jungong, Debattista, Kurt,1975- | ||||
Sponsors: | Southern University of Science and Technology ; University of Warwick | ||||
Extent: | xvi, 166 pages : illustrations (some colour, some black and white) | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |