In this episode, we talk with Heather Couture about how to make deep learning models for tissue image analysis more robust to domain shift.
Supervised deep learning has made a strong mark in the histopathology image analysis space, however, this is a data-centric approach. We train the image analysis solution on whole slide images and want them to perform on other whole slide images – images we did not train on.
The assumption is that the new images will be similar to the ones we train the image analysis solution on, but how similar do they need to be? And what is domain and domain shift?
Domain: a group of similar whole slide images (WSI). E.g., WSIs coming from the same scanner or coming from the same lab. We train our deep learning model on these WSIs, so we call it our source domain. We later want to use this model and target a different group of images, e.g. images from a different scanner or a different lab – our target domain.
When applying a model trained on a source domain to a target domain we shift the domain and the domain shift can have consequences for the model performance. Because of the differences in the images the model usually performs worse…
How can we prevent it or minimize the damage?
Listen to Heather explain the following 5 ways to handle the domain shift:
- Standardize the appearance of your images with stain normalization techniques
- Color augmentation during training to take advantage of variations in staining
- Domain adversarial training to learn domain-invariant features
- Adapt the model at test time to handle the new image distribution
- Finetune the model on the target domain