In this PhD, the goal will be to research advanced Geometric Deep Learning algorithms for the discovery of novel spatial biomarkers in gigapixel whole slide images in histopathology.To date, most algorithms in Geometric Deep Learning assume graphs are already existing and known, and message passing helps to learn the features that are communicated between locations in the graph. An open question, therefore, is whether one can assume the graph is not known and must be discovered instead, so that it is a) consistently recognisable across samples, and b) descriptive enough for the final predictive task at hand.
In histopathology, images are basically digital scans of tissue biopsies in the range of gigapixel sizes, comprising abstract patterns of biological materials like cells, vessels, fat, tissue, and of course tumours and antibodies. In trying to understand whether a highly-specialized treatment, such as the Nobel-winning immunotherapy, is a good fit, specific biomarkers must be recognized in the histopathological whole slide image. Unfortunately, due to the complexity and the size of the whole slide images, current biomarkers mostly focus on the presence or absence of specific cell types, and their densities. A strong hypothesis is that not only the presence or absence of cell types is important, but also the way they are organized spatially over the patient's tissue. While this is complex to systematically infer with the naked eye, and routinely recognize in clinical practice, the hope is that Deep Learning algorithms that rely on Graphs and Geometry can be of great complemenatry value.
The project is between the University of Amsterdam, the Amsterdam University Medical Center, and Ellogon.AI.
medical backgroundThe project focuses specifically on
Gastro-Esophageal cancer (GEC). Novel immunotherapy treatment (IMT) shows promising therapeutic results in a subset of all cancer patients but is highly costly.
To select patients that will benefit, medical specialists like pathologists need to quantify biomarkers that are predictive for IMT outcome. Various cancer types require either PD-L1, or tumor infiltrating lymphocytes (TIL) or tumor-foreignness biomarker testing. These biomarker assessments are however complex, expensive and suffer from interobserver variability. Current limited testing methods result in unwanted variation in patient outcome and high healthcare costs. An evident clinical need exists for objective, integrated and easy to use decision support tools, to optimize personalized treatment and identify GEC patients that will respond to IMT.
We will address these needs by developing an integral system for biomarker quantification directly from standard histopathology H&E slides without the need of performing additional test. Based on probabilistic and geometric deep learning principles, we intend to model multi-scale, spatial, and contextual dependencies through geometric (hyper)graph neural networks. Spatio-contextual cues, such as the presence of lymphocytes within cancer tissue, form biomarkers that the system is to predict.
Subsequently, we aim to reliably predict the outcome of IMT for individual patients, based on the observed and quantified biomarkers. In order to maintain transparency and understanding of how the biomarkers relate to predicted outcome, we rely on advanced probabilistic modeling. Throughout the project, we will use state-of-the-art deep learning approaches such as self-supervised learning to make most use of all available data and group equivariant learning approaches to improve data efficiency.
Our goal is to bring AI solutions into the daily clinical practice of pathologists and oncologists to identify patients who may benefit most from immunotherapy or could be spared unnecessary treatment. Embedded in the thriving Amsterdam AI ecosystem, you will develop cutting-edge AI solutions at a fundamental level, whilst working at the forefront of IMT research.
About your roleAs a PhD-candidate, you will be responsible for developing and evaluating state-of-the-art deep learning techniques in multidisciplinary medical data. You will be involved in preparing histopathology datasets of GEC patients for cancer-immune interaction and clinical outcome data. Finally, you will validate the algorithms you developed with immunotherapy treatment outcome in independent patient cohorts to ensure the devised AI-algorithms' applicability in clinical practice.
- You will collaborate with other researchers within the research labs of the SELECT-AI consortium (Amsterdam UMC departments of Pathology and Medical Oncology, University of Amsterdam the Institute of Informatics, and the department of Pathology from NKI-AvL).
- Regularly present internally on your progress
- Regularly present intermediate research at international conferences and workshops, publish them in proceedings and journals, help with submitting applications
- Assist in relevant teaching activities
- Complete and defend a PhD thesis