An analytical framework of tissue-patch clustering for quantifying phenotypes of whole slide images

Sunhem, Wisuwat (2023) An analytical framework of tissue-patch clustering for quantifying phenotypes of whole slide images. MSc(R) thesis, University of Glasgow.

Abstract

Histopathology is considered the most practical diagnostic method for patient with early stage cancer. This is because at the very first pre-screening, patient’s tissue samples are delivered to pathologist for examining evidence of cancer. Computational scientists aid pathologist by heavily producing research on machine learning-based morphological pattern recognition of tissue image. Many data modelling investigations on histopathology have been conducted in supervised manner and some of them were further employed in real-life clinical diagnosis. This study proposes an approach to developing clusters of tissue tile. The main aim is to obtain ’high-quality clusters’ with respect to phenotypic annotations. In order to achieve this goal, two colorectal datasets namely 100k-nct and TCGA-COAD are experimented, one of which is directly annotated with tissue type, and other dataset is annotated through derivation from patient metadata, quiescent status. Four main independent variables were explored in this study (i) feature extraction by Resnet50, InceptionV3, VGG16 and an unsupervised generative model, PathologyGAN. (ii) feature space transformer including original feature, 3D PCA feature and 3D-UMAP feature and (iii) clustering algorithms namely Gaussian Mixture Model and Hierarchical clustering and their primary hyper-parameters. As a result, Resnet50 empowered by UMAP outperformed the most in clustering tissue type on 100k-nct dataset at v-measure of 0.74. The other dataset of which quiescent status is derived from patients encountered nearly zero in v-measure. However, clustering this quiescence-based dataset on 3D-UMAP Pathology-GAN yielded far higher V-measure than the rest of cluster configurations and illustrates ability to capture quiescence-related phenotype through visualisation.

• Item Type: Thesis (MSc(R))
• Qualification Level: Masters
• Keywords: Phenotype cluster, deep learning, generative adversarial model, tumour tissue, manifold learning, dimension reduction.
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science; Q Science > QA Mathematics > QA76 Computer software; R Medicine > RC Internal medicine > RC0254 Neoplasms. Tumors. Oncology (including Cancer)
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Yuan, Dr. Ke and Jensen, Dr. Bjorn Sand
• Date of Award: 2023
• Depositing User: Theses Team
• Unique ID: glathesis:2023-83451
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 23 Feb 2023 12:21
• Last Modified: 28 Feb 2023 09:02
• Thesis DOI: 10.5525/gla.thesis.83451
• URI: https://theses.gla.ac.uk/id/eprint/83451