Developing a self-supervised learning pipeline to capture low dimensional representations of placenta histology to phenotype pathological changes

This project aims to develop novel computational methods for analysing placental histology images to better understand maternal health outcomes. The placenta plays a crucial role in fetal development and can provide valuable insights into both fetal and maternal health. However, placental analysis remains challenging due to the organ's spatial and temporal complexity and heterogeneity. The research will leverage deep learning and computer vision techniques to extract biologically meaningful information from whole slide images (WSIs) of placental tissue.
Specifically, the project will extend an existing supervised 3-stage pipeline (nuclei-cell-tissue) to a 4-stage self-supervised learning pipeline (nuclei-cell-tissue-pathology) to generate low- dimensional representations of placental histology that capture features without relying on expert annotations. This approach has the potential to reveal patterns and phenotypes that may not be apparent through traditional analysis methods and will allow for high- throughput analysis of placental tissue. Downstream this can be used to characterise placenta pathologies and understand how placenta variation affects maternal health.
The key objectives of the project are:
1. Develop a 4-stage self-supervised representation learning pipeline for placental
histology images, extending existing work on nuclei, cell, and tissue classification in WSIs to include pathology identification, enabling characterisation of normal variation and pathological changes of placental tissue.
2. Develop methods for graph-based representation learning to compress and interpret WSI data to enable downstream analysis to cluster placentas representations.
3. Investigate the clinical relevance of placental histology representations for predicting maternal health outcomes using longitudinal cohort data.
The novelty of the research methodology is in several aspects. This project will develop self- supervised learning techniques to the understudied domain of placental pathology. By avoiding reliance on expert annotations, the methods developed have the potential to reveal unbiased biological patterns. The project introduces a multistage interpretable pipeline, where each stage of characterising the placenta (nuclei-cell-tissue-pathology) can be understood and analysed independently, with potential applications for a foundation histology model. The project will develop a new graph autoencoder architecture, designed to compress the WSI information to enable efficient downstream analysis while preserving structural information.
Overall, the project has the potential for significant impact by automating placental analysis capabilities to extract biologically meaningful information. This could lead to improved understanding of placental pathologies and their relationship to maternal health and post- partum risks.
This project falls within the EPSRC Healthcare Technologies theme, specifically the 'Image and Vision Computing' and 'Biological Informatics' research areas.