Project Description
Together with colleagues at the Biomedical Sciences department at Radboudumc, we run an imaging platform for cancer research, based on the so-called "Vectra Polaris" digital microscopy platform. To understand the background of this system, you can watch a brief promotional video (in Dutch, with the popular singer “Do”) about the underlying research here: https://www.youtube.com/watch?v=VHRHWJw6jcA and a more technical video that explains the system is available here: https://app.jove.com/v/65717/author-spotlight-unlocking-insights-into-immune-cell-landscape
Over the years, we have collected several terabytes of human tissue sample images and we would now like to leverage this treasure trove of data to build smarter biomedical computer vision systems for the future. Specifically, when training our existing machine learning pipeline to detect and classify cells in these images (https://doi.org/10.1093/biomethods/bpae094), we have created a hand-curated dataset of almost 200,000 cell annotations. We would now like to use a self-supervised vision transformer model -- perhaps DinoV3 by Meta -- to automatically scale up this annotation dataset by approximately 10-fold. We would then like to use this "10xed" dataset to determine whether it leads to performance improvements for our existing machine learning models or perhaps even enables us to train our own foundation model that is specific for this kind of data.
People
Johannes Textor
Requirements
Overview