Computational insights into mouse placental histology

Thesis

This thesis represents a significant contribution to the field of placental biology, combining advanced computational techniques with traditional biological research, with a focus on computer vision methods. I utilised the Deciphering the Mechanisms of Developmental Disorders (DMDD) dataset as a factual basis, confirming and demonstrating the capabilities of computational paradigms in the detailed analysis of placental histology.

The research consists of four distinct phases, each aiming to provide in-depth information for the investigation. The first phase involved thorough annotation and preparation of the DMDD dataset to ensure that subsequent computational analyses were methodologically sound and valid. This phase formed the foundation for the rest of the study.

The second phase introduced an automated placental phenotyping framework that uses machine learning algorithms such as RetinaNet and InceptionResNetV2. This approach transcended species boundaries, encompassing both mouse and human placentae, and improved the categorisation of cellular and histological features with state of the art accuracy. This framework represented a significant shift in approach by eliminating the limitations of traditional manual labor and reducing interpretive subjectivity.

During the third phase of the study, the DMDD dataset was thoroughly analyzed using an automated pipeline. This analysis helped to identify unknown phenotypic markers, particularly in E14.5 mouse placentae. The newfound markers were carefully verified through scholarly corroboration and cross-referencing, and they played a crucial role in advancing knowledge of the complex interplay between placental and embryonic phenotypes in knockout mice. Machine learning models like XGBoost were used for further feature analysis to support the findings, adding an extra layer of empirical robustness.

In the last phase of this thesis, I explored the intricate world of unsupervised and self-supervised learning techniques. I made use of advanced computational approaches such as t-distributed Stochastic Neighbour Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and SimCLR to gain a deeper understanding of the intraclass variance in the cellular structure of placental tissue. My investigations provided initial insights into this complex field and highlighted methodological guidelines, challenges, and opportunities for future research.

This thesis builds on existing literature as a catalyst for transforming thinking in the field. The study demonstrates how computational methods, specifically computer vision, can revolutionize our comprehension of placental biology and developmental studies. This research blends technological innovation with biological discovery to enhance our understanding of the placenta.

Authors: McCann Strain, P

• DOI: 10.5287/ora-mvz4vx2g8
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning algorithms; computational biology; phenotypic marker discovery; knockout mice phenotypes; placental biology; unsupervised learnin; deep learning; automated placental phenotyping; deciphering the mechanisms of developmental disorders (dmdd) dataset; computer vision methods; placental histology
• Subjects: Tissue Imaging; Automated Phenotyping; Knockout Mice Studies; Computer Vision; Genetics; Bioinformatics; Histology; Developmental Biology; Computational Biology; Developmental Disorders; Placental Biology; Machine Learning; Deep Learning; Phenotypic Marker Discovery