Modelling the placenta from the bottom up: digital phenotyping of human placenta histology

Thesis

Accurate placenta histopathology assessment is essential for immediate and lifelong clinical management of mother and newborn health. However, the placenta’s heterogeneity and high disease tolerance are challenging for robust, reproducible histological analysis and detection of pathological changes. Placenta histology slides contain upwards of a million cells and tens of thousands of micro-anatomical tissues. High-throughput, quantitative and objective metrics of placental biology are therefore valuable for placental investigations in clinical and research settings. Deep learning for digital pathology has the potential to provide these high-throughput metrics, but the placenta has received relatively little attention compared to other organ histology.

In this work, I present a three-stage hierarchical deep learning pipeline for analysing the placenta from the bottom up. The pipeline outputs rich and interpretable biological metrics following the organ’s anatomical hierarchy. I start the pipeline at the cellular level, localising all nuclei across a slide and classifying them into one of 11 cell types. Using the nuclei coordinates and cell classes, I construct a whole slide cell graph that mirrors cellular community interaction within micro-anatomical tissue structures. I develop a scalable graph neural network which uses constituent cells to predict one of 9 tissue microstructures in the placenta parenchyma.

The whole slide cellular and tissue microstructure predictions match expectations from placental biology literature and closely replicate those from independent clinical experts. I use these whole slide metrics to quantify healthy variation in the term placenta and find that significant pathological changes are caused by common placental lesions. Finally, I develop a novel graph compression autoencoder with a custom compression algorithm for unsupervised region clustering. This work is a step towards automated, interpretable and high-throughput metrics for assessing placental health, with potential applications in both clinical and research settings.

Authors: Vanea, C

• DOI: 10.5287/ora-mz2y41zdn
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: computer vision; histology; digital pathology; graph neural networks; placenta; AI
• Subjects: Placenta; Histology; Deep learning (Machine learning)