Predictive Biomarkers from Multiplex Immunofluorescence

Problem: While H&E is the most common modality for studying tissue, it only highlights two things: nuclei and cytoplasm. Multiplex immunofluorescence (mIF), on the other hand, characterizes many different proteins and cell types simultaneously. This more complete picture of tissue biology can enable more powerful biomarkers for predicting patient outcomes. However, these additional channels can bring new challenges to training a deep learning model.

Solution: This project began with providing guidance on nuclei segmentation approaches for multiplex imaging. Once individual cell types could be identified, we focused on image translation models that could predict expression levels from a subset of channels. These were used to predict cell types and assess how well different cell types could be predicted.