Automatic Phenotyping of Microscopic Skin Images

Hussein, Saif (2019) Automatic Phenotyping of Microscopic Skin Images. Doctoral thesis, University of Buckingham.

Saif Raoof Hussein_Automatic Phenotyping of Microscopic Skin_PhDThesis2020.pdf
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (9MB) | Preview


The Mouse Genetics Project (MGP) is a large-scale mutant mice production and phenotyping initiative that builds on the success of the Human Genome Project (HGP) to discover the functionality of all genes and their role in human diseases. The MGP aims to produce over 20,000 mutant lines of mouse model to investigate and quantify the impact of gene knockouts on the various organs. Due to the great deal of overlap between human and mouse genomes, most the acquired knowledge can be translated into diagnostics, biomarker identifications and eventually treatment for complex genetic-based diseases. The skin is by far the largest organ in the mammalian body, and a complex structure of multiple layers consisting of different distinguishable cells and objects. Dermatology specialists have long associated many diseases with changes in different skin layers such as changes to the number of nuclei in the dermis and epidermis layers, the orientation of hair follicles, curvature of the outer border of the epidermis, and many more. The manual quantification and analysis of such features/objects for a high throughput phenotype research project such as the MGP is an error-prone, time consuming and very costly in terms of resources and staff training. Recent rapid advances in biomedical image processing/analysis as well as the emergence of a variety of machine learning tools provide an exciting motivation to developing effective and efficient automatic solutions. Automation is as challenging as the manual methods but the challenges are of different nature. Therefore, this thesis is devoted to investigate, develop and test a number of automatic algorithms to quantify the above mentioned features/objects in mouse skin layers and experimentally identify the genetic causes of changes to these parameters in relation to skin diseases. Our investigations and solutions had to deal with a number of technical challenges such as staining errors that lead colour overlapping between neighbouring layers/components, damages to the outer layers during preparation of images, the difficulty of establishing the ground truth for the large volume of image dataset, significant overlapping of nuclei objects, misalignments of tissues with the large volume of dataset. The main contributions of the thesis include: 1. Proposing an adaptive system that combines colour deconvolution and fuzzy c-mean methods to segment the three main skin layers in H&E images, namely the epidermis, dermis and fat cell layers, after establishing the limitation of existing solutions to overcome the challenges highlighted above. 2. Developing automatic methods for segmenting and counting nuclei in the epidermis and dermis layer in mice skin and demonstrating the ability of the proposal in identifying overlapping nuclei and separating them. Furthermore, we automatically identify a list of candidate genes responsible for abnormal changes to the number of nuclei. 3. Introducing a simple method to align the epidermis outer border in all images and designing an easy geometric-based formula to quantify the orientation of hair follicles with respect to the aligned epidermis as an indicator of skin abnormality. This led to identifying a list of candidate genes responsible for abnormal changes to the orientation of hair follicles. 4. Defining a simple mathematical model of the vaguely defined epidermis curvature as an indicator of skin abnormality and proposing a reliable scheme to quantify it. 5. Providing empirical evidences of the success for each of the above schemes by comparing the automatically determined quantification outputs with the ground truth determined by domain expert biology researchers. 6. Demonstrating the high potential of non-invasive, unsupervised machine learning techniques in the successful isolation of potentially interesting (knockout genes) relevant to genetic causes of skin abnormalities. This would facilitate high-throughput analysis in cutaneous research, with potential applications for screening drugs.

Item Type: Thesis (Doctoral)
Uncontrolled Keywords: mouse genetics project; MGP; phenotyping; biomedical image processing; dermis layers; epidermis layers; colour deconvolution; machine learning; epidermis curvature
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Q Science > QA Mathematics > QA76 Computer software
Divisions: School of Computing
Depositing User: Rachel Pollard
Date Deposited: 26 Nov 2021 11:40
Last Modified: 26 Nov 2021 11:40

Actions (login required)

View Item View Item