The Library
Working with scarce annotations in computational pathology
Tools
Koohbanani, Navid Alemi (2020) Working with scarce annotations in computational pathology. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Koohbanani_2020.pdf - Submitted Version - Requires a PDF viewer. Download (15Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3519986~S15
Abstract
Computational pathology is the study of algorithms and approaches that facilitate the process of diagnosis and prognosis of primarily from digital pathology. The automated methods presented in computational pathology decrease the inter and intra-observability in diagnosis and make the workflow of pathologists more efficient. Digital slide scanners have enabled the digitization of tissue slides and generating whole slide images (WSIs), allowing them to be viewed on a computer screen rather than through a microscope. Digital pathology images present an opportunity for development of new algorithms to automatically analyse the tissue characteristics.
In this thesis, we first focus on the development of automated approaches for detection and segmentation of nuclei. In this regard, for nuclear detection, each nucleus is considered as a Gaussian shape where the mean of Gaussian determines the centroids of nuclei. We investigate the application of mixture density networks for detection of nuclei in the histology images.
We also propose a convolutional neural network (CNN) for instance seg mentation of nuclei. The CNN uses the nuclei spatial information as the target to separate the clustered nuclei. Pixels of each nucleus are replaced with the spatial information of that nucleus. The CNN also utilises dense blocks to reduce number of parameters and positional information at different layer of the network to better learn the spatial information embedded in ground truth.
Two chapters of this thesis are dedicated to dealing with lack of annotations in computational pathology. To this end, we propose a method named as NuClick to generate high quality segmentations for glands and nuclei. NuClick is an interactive CNN based method, that requires minimum user interaction for collecting annotations. We show that one click inside a nucleus can be enough to delineate its boundaries. Moreover, for glands that are more complex and larger objects a squiggle can extract their precise outline.
In another chapter, we propose Self-Path, a method for semi-supervised learning and domain alignment. The main contribution of this chapter is proposing self-supervised tasks that are specific to histology domain and can be extremely helpful when there are not enough annotations for training deep models. One of these self-supervised tasks is predicting the magnification puzzle which is the first domain specific self-supervised task shown to be helpful for domain alignment and semi-supervised learning for classification of histology images.
Nuclear localization allows further exploration of digital biomarkers and can serve as a fundamental route to predicting patient outcome. In chapter 6, by focusing on the challenge of weak labels for whole slide images (WSIs) and also utilising the nuclear localisation techniques, we explore the morphological features from patches that are selected by the model and we observe that these features are associated with patient survival.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software R Medicine > RB Pathology T Technology > TA Engineering (General). Civil engineering (General) |
||||
Library of Congress Subject Headings (LCSH): | Image processing -- Digital techniques, Pathology -- Data processing, Histology, Pathological, Neural networks (Computer science), Cell nuclei -- Counting | ||||
Official Date: | November 2020 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Department of Computer Science | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Rajpoot, Nasir M. (Nasir Mahmood) | ||||
Sponsors: | Alan Turing Institute ; Intel Corporation | ||||
Format of File: | |||||
Extent: | xix, 145 leaves : illustrations (chiefly colour) | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |
Downloads
Downloads per month over past year