The Library
Learning with minimal annotations in computational pathology
Tools
Bashir, Raja Muhammad Saad (2023) Learning with minimal annotations in computational pathology. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Bashir_2023.pdf - Unspecified Version - Requires a PDF viewer. Download (104Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3977180~S1
Abstract
Deep learning has pushed the boundaries of Computational Pathology (CPath) models for the diagnosis and prognosis of cancer. Many methods have been proposed that are fast, reliable and reproducible, but the performance largely depends on large scale labelled data. In most cases, a large amount of data remains unlabelled and needs to be used. Therefore, this thesis focuses on developing semi-supervised and weakly-supervised approaches for automated analysis of whole slide images (WSIs) leveraging unlabelled data. To this effect, I present a semi-supervised method for simultaneously classi fying and detecting tumour cells in Diffuse Large B-Cell Lymphoma (DLBCL). I first label the unlabelled data using pseudo labels and then train the frame work using MixUp augmentation, which enhances the generalisation capability of the network. Next, I segment nuclei and tissue regions in WSIs using semi supervised and self-supervised learning. Limited labelled data challenges the model’s robustness due to limited exposure and learning experience. Therefore, I propose a consistency regularisation and cross-consistency training based semi-supervised learning framework. In addition, I also incorporate entropy minimisation to improve the confidence of pseudo labels predicted during training. Finally, I use multiple instance learning (MIL) frameworks for the diagnosis (i.e., grading) and prognosis (i.e., malignant transformation) of oral epithelial dysplasia (OED). I propose a novel digital biomarker, based on a count of peri epithelium lymphocytes, and demonstrate its association with poor progression free survival (PFS) in OED. Then, I propose a method based on graph neural networks (GNN) in a larger cohort. Initially, I perform coarse segmentation to delineate the epithelium into sub-layers and then train GNN models with ranking loss. The findings reveal that nuclei from the epithelium and basal layers are significant diagnostic digital biomarkers for grading. In contrast nuclei from the basal layer and peri-epithelium tissue area are found to be significant for OED malignant transformation.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > QA Mathematics > QA75 (Please use QA76 Electronic Computers. Computer Science) R Medicine > RC Internal medicine > RC0254 Neoplasms. Tumors. Oncology (including Cancer) |
||||
Library of Congress Subject Headings (LCSH): | Deep learning (Machine learning), Cancer -- Diagnosis|xData processing, Diagnostic imaging -- Data processing, Pathology -- Data processing, Artificial intelligence -- Medical applications, Computer vision | ||||
Official Date: | June 2023 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Department of Computer Science | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Rajpoot, Nasir M.|q(Nasir Mahmood) ; Raza, Shan-e-Ahmed | ||||
Sponsors: | University of Warwick.Chancellor's International Scholarship | ||||
Extent: | xxv, 149 leaves : illustrations | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |