The adoption of digital pathology has enabled the curation of large repositories of gigapixel whole-slide images (WSIs). WSIs are invaluable for examining cellular morphology and how it changes over embryonic development or disease progression. Many existing methods employ well-trained deep neural networks to extract image features from histology images, and then use these image features for downstream analysis. A drawback of using deep neural networks, e.g., ResNet and Vision Transformer (ViT), is that these models require a large number of well-annotated images from pathologists for model training, which limits their usefulness. This project aims to develop label-free machine learning methods for medical imaging data analysis. The developed tools can be easily applied to studies where training samples are not available and bypass cumbersome labeling steps.
In collaboration with: Eliot Graham Peyster, Nan Ma